Kvalitnejšie odpovede

README.md

@@ -1160,7 +1160,7 @@ upozornenie na neúplnosť, nezamieňať autora dokumentu s autorom práce a pri
 sa nevysvetľujú odhadom; verzia dát neoznačuje verziu programu.
 Tieto pokyny ovplyvňujú správanie Claude, jeho finálny text však náš backend
 nemôže programovo vynútiť. Po aktualizácii už nahratých Skills nahraď ich
-novými ZIPmi z `packages/claude/dist/` (verzia 1.2.0) a over správanie v novom chate.
+novými ZIPmi z `packages/claude/dist/` (verzia 1.3.0) a over správanie v novom chate.
 
 ### Nahratie Skills do webového Claude
 
@@ -1507,6 +1507,22 @@ Chýbajúce výsledky sa oznámia, neodvodzujú sa z cieľov. Grafová otázka n
 v ktorých majú práce „obaja“, používa prienik rokov a prikladá doložené typy a názvy
 prác z týchto rokov; nezobrazuje namiesto odpovede všetky individuálne roky.
 
+`rag_preparation.py` vie pomenované profily dohľadať v SQLite aj mimo prvých
+výsledkov a pri viacerých možných prácach vyžiada typ alebo rok. Obsahové
+porovnanie je obmedzené na štyri osoby naraz. `question_intent.py` rozlišuje
+požadované problémy, metódy, experimenty, ciele a výsledky. `content_selection.py`
+vyberá doslovné, typovo a osobne viazané podklady, odstraňuje opakovania a
+uprednostňuje vecné výsledky pred administratívnymi poznámkami. Model dostane
+len tieto kandidátske výňatky; ich skutočný JSON podlieha tokenovému rozpočtu.
+Chýbajúce časti sa označia jednotlivo, aj keď sú ostatné časti zodpovedateľné.
+Pri limite kontextu sa nestratí informácia, ktoré osoby používateľ požadoval.
+
+Rozpoznané, ale nepodporované operátory (napríklad interval rokov, vylučovanie
+alebo minimálny počet) vedú k `ambiguous` a konkrétnemu spresneniu. Rok práce
+sa nepoužije ako dôkaz dátumu jej začatia či obhajoby. Chybové odpovede odlišujú
+nedostupné vyhľadávanie, nedostupný model, vyťaženie a zamietnutý neoverený text;
+raw výstupy odmietnutého modelu sa štandardnému klientovi neposkytujú.
+
 Otázky na obsah, zadanie, ciele alebo metódy práce používajú textové výňatky
 z príslušnej akademickej sekcie, nie iba metadáta o názve a roku. Výňatok musí
 patriť požadovanej osobe, typu práce a prípadne roku. Názov, rok ani samotný

app/routes.py

@@ -49,6 +49,7 @@ from tools.retrieval.publication import DEFAULT_PUBLISHED_ONLY
 from tools.runtime.tool_health import tools_readiness
 from tools.retrieval.retrieval_contract import RetrievalResponse, RetrievalSource, RetrievalErrorResponse, error_reason
 from tools.answers.answer_delivery import deliver_answer
+from tools.answers.answer_errors import public_failure
 
 
 logger = logging.getLogger(__name__)
@@ -192,9 +193,7 @@ def verified_answer(request: AnswerRequest) -> dict:
                 'answer': 'Overenú odpoveď sa nepodarilo pripraviť. Skúste to neskôr.'}
     if (result.get('accepted_for_delivery') is not True or
             result.get('status') not in {'ok', 'no_answer', 'ambiguous'}):
-        result = {**result, 'status': 'error', 'accepted_for_delivery': False,
-                  'answer': 'Odpoveď neprešla overením podkladov. Neoverené údaje sa nezobrazia.',
-                  'claims': [], 'source_urls': []}
+        result = public_failure(result)
     if not request.include_diagnostics:
         result.pop('attempts', None)
         result.pop('retrieval_context', None)

packages/claude/client.py

@@ -81,8 +81,22 @@ class Config:
 
 
 def error_result(reason):
+    messages={
+        'unauthorized':'Prístup nebol autorizovaný. Skontroluj prihlásenie konektora a jeho konfiguráciu.',
+        'forbidden':'Pre tento nástroj nemáš potrebné oprávnenie.',
+        'invalid_query':'Otázku sa nepodarilo spracovať. Zadaj konkrétnu otázku a primeraný počet výsledkov.',
+        'model_unavailable':'Jazykový model momentálne nie je dostupný. Skús otázku neskôr.',
+        'answer_rejected':'Odpoveď sa nepodarilo spoľahlivo overiť podľa podkladov. '
+                          'Spresni osobu, typ práce alebo rok; neoverený text sa nezobrazí.',
+        'answer_capacity_exceeded':'Agent práve spracúva iné otázky. Skús to o chvíľu.',
+        'backend_busy':'Agent práve spracúva iné otázky. Skús to o chvíľu.',
+        'backend_timeout':'Vyhľadávanie neodpovedalo včas. Skús otázku zopakovať neskôr.',
+        'backend_unavailable':'Spojenie s backendom momentálne nie je dostupné. Skontroluj, či bežia služby a tunel.',
+        'retrieval_unavailable':'Vyhľadávací index momentálne nie je dostupný. Skús otázku neskôr.',
+    }
     return {'status': 'error', 'reason': reason, 'accepted_for_delivery': False,
-            'answer': '', 'sources': [], 'index_version': None, 'complete': False,
+            'answer': messages.get(reason,'Overenú odpoveď sa momentálne nepodarilo získať.'),
+            'sources': [], 'index_version': None, 'complete': False,
             'truncated': False, 'truncation_reasons': []}
 
 

packages/claude/dist/packages.json

@@ -3,12 +3,12 @@
     {
       "archive": "zpwiki-rag.zip",
       "name": "rag",
-      "sha256": "8f0c51e27dc68a1e3ee845f2cfe5987e51714d4408f98c7d275861dc1e22953f"
+      "sha256": "09d3e6b196b61f6c025eb86ccd1773b913b37660163a995ad1a2acfa6d947a0c"
     },
     {
       "archive": "zpwiki-graphrag.zip",
       "name": "graphrag",
-      "sha256": "aa65a97211f48fa0352a98e50c6ca6a243e1d23105c0ff144c7f360ad192c376"
+      "sha256": "cfcb60213e103d40c2adbc6c7d4133751120148415836cfffa7935c020167e95"
     }
   ],
   "schema_version": 1

scripts/build_claude_packages.py

@@ -11,7 +11,7 @@ import zipfile
 ROOT = Path(__file__).resolve().parents[1]
 BASE = ROOT / 'packages/claude'
 ENGINES = ('rag', 'graphrag')
-PACKAGE_VERSION = '1.2.0'
+PACKAGE_VERSION = '1.3.0'
 
 
 def json_bytes(value):

test/test_claude_packages.py

@@ -156,7 +156,7 @@ def test_invalid_answer_coverage_is_not_delivered(backend, coverage):
 def test_transport_errors_and_redirects_do_not_disclose_payload(backend, code, reason):
     backend.update(raw=b'SECRET ERROR', code=code)
     value = client.query_backend('Práca?', 'rag')
-    assert value['reason'] == reason and value['answer'] == ''
+    assert value['reason'] == reason and value['answer']
     assert len(backend['requests']) == 1 and 'SECRET' not in json.dumps(value)
 
 
@@ -168,7 +168,7 @@ def test_rejected_or_malformed_answer_is_never_delivered(backend, change):
     backend.update(raw=json.dumps({**backend['checked'], **change}).encode())
     value = client.query_backend('Práca?', 'rag')
     assert value['status'] == 'error' and not value['accepted_for_delivery']
-    assert value['answer'] == '' and 'Eva' not in json.dumps(value)
+    assert value['answer'] and 'Eva' not in json.dumps(value)
 
 
 def test_invalid_json_and_oversized_response(backend):
@@ -279,7 +279,7 @@ def test_mcp_stdio_initialization_discovery_call_and_rejection(backend):
                     assert all('presentation_instructions' not in json.loads(block.text)
                                for block in result.content if block.type == 'text')
                 rejected = await session.call_tool('zpwiki_rag', {'query': ''})
-                assert rejected.isError and rejected.structuredContent['answer'] == ''
+                assert rejected.isError and rejected.structuredContent['answer']
     asyncio.run(run())
 
 

test/test_comparison_answers.py

@@ -103,8 +103,11 @@ def comparison_payload(tmp_path):
 
 
 def selected_comparison(context):
-    quotes=[{'source_id':p['source_id'],'quote':s['text']} for p in context['content_passages']
+    quotes=([{'source_id':c['source_id'],'quote':c['text']} for c in context['content_candidates']]
+            if 'content_candidates' in context else
+            [{'source_id':p['source_id'],'quote':s['text']} for p in context['content_passages']
             for s in content_sections(p['text']) if s['kind'] in {'assignment','progress'}]
+            )
     return {'status':'ok','answer':'','claims':[],'evidence_quotes':quotes,
             'source_urls':[s['source_url'] for s in context['sources']]}
 
@@ -169,7 +172,7 @@ def test_missing_results_allow_an_explicit_partial_comparison(comparison_payload
     selected=selected_comparison(c)
     result=generate_verified_rag(COMPARE,payload,lambda _: {'choices':[{'message':{'content':json.dumps(selected)}}]})
     assert result['accepted_for_delivery']
-    assert 'nie sú dosiahnuté výsledky doložené' in result['answer']
+    assert 'Táto časť odpovede nie je v poskytnutých podkladoch doložená.' in result['answer']
     assert 'Implementácia systému vyhľadávania' in result['answer']
 
 
@@ -182,9 +185,9 @@ def test_future_tasks_are_not_presented_as_achieved_results(comparison_payload):
     p=next(p for p in c['content_passages'] if p['person']=='Ján Holp')
     selected['evidence_quotes'].append({'source_id':p['source_id'],'quote':'- Pripravte experiment s PageRank.'})
     result=generate_verified_rag(COMPARE,payload,lambda _: {'choices':[{'message':{'content':json.dumps(selected)}}]})
-    assert result['accepted_for_delivery']
-    assert '**Ďalšie doložené úlohy (plán):**' in result['answer']
-    assert 'Pripraviť experiment s PageRank.' in result['answer']
+    # The model may not quote a future task omitted from its actual candidates.
+    assert not result['accepted_for_delivery']
+    assert 'quote_outside_delivered_candidates' in result['attempts'][-1]['issues']
 
 
 @pytest.fixture

test/test_quality_block.py

@@ -0,0 +1,262 @@
+"""Behavioral checks for scope, relevant evidence, partial answers and failures."""
+import copy
+import json
+import sqlite3
+
+import pytest
+
+from tools.answer_evaluation.content_selection import candidates, facets_for
+from tools.answer_evaluation.rag_evidence import rag_evidence_context
+from tools.answer_evaluation.rag_preparation import prepare_named_content
+from tools.answer_evaluation.verified_rag import generate_verified_rag, rag_answer_issues
+from tools.answers.answer_errors import public_failure
+from tools.graphrag import graph_search as gs
+from tools.graphrag.tests.test_graph_search import FakeSession
+from tools.rag.llm_context import compact_rag_payload
+from tools.rag.rag_utils import build_context_text
+from tools.retrieval.question_constraints import unsupported_constraint, requested_years
+
+
+@pytest.fixture
+def corpus(tmp_path):
+    db=tmp_path/'index.sqlite'; sources=[]
+    with sqlite3.connect(db) as conn:
+        conn.execute('CREATE TABLE documents(path,title,author,published)')
+        conn.execute('CREATE TABLE chunks(document_path,text,heading_paths_json,published,chunk_index)')
+        for i,(person,slug,kind,year,text) in enumerate([
+            ('Eva Nová','eva_nova','Diplomová práca',2024,
+             'Zadanie:\n1. Implementujte systém vyhľadávania.\nStav:\n- Konzultácia štruktúry práce.\n'
+             '- Natrénovaný model BERT a vykonané vyhodnotenie presnosti.\n'
+             '- Vznikol problém s tokenizérom.\nÚlohy:\n- Pripravte experiment s rôznymi tokenizérmi.'),
+            ('Boris Nový','boris_novy','Bakalárska práca',2023,
+             'Zadanie:\n- Implementujte parser.\nStav:\n- Implementovaný parser pomocou Python.\n'
+             '- Vznikla chyba pri načítaní PDF.\nÚlohy:\n- Pripravte experiment s presnosťou extrakcie.')],1):
+            path=f'pages/students/2020/{slug}/README.md'
+            conn.execute('INSERT INTO documents VALUES(?,?,?,?)',(path,person,'Daniel Hladek',1))
+            conn.execute('INSERT INTO chunks VALUES(?,?,?,?,?)',(path,text,json.dumps([[person,f'{kind} {year}']]),1,0))
+            sources.append({'source_id':f'S{i}','title':person,'document_path':path,'published':True,
+                            'source_url':f'https://example.test/{slug}',
+                            'text':f'Sekcia: {kind} {year}\nNázov: Ukážkový názov.'})
+    raw={'status':'ok','reason':None,'sources':sources,'context':build_context_text(sources),
+         'complete':False,'truncated':True,'truncation_reasons':['result_limit'],'index_version':{'version':'fixed'}}
+    return db,raw
+
+
+def prepare(corpus,question):
+    db,raw=corpus
+    return compact_rag_payload(prepare_named_content(db,question,raw))
+
+
+def model_selection(question,payload):
+    c=rag_evidence_context(question,payload)
+    return {'status':'ok','answer':'','claims':[],
+            'evidence_quotes':[{'source_id':x['source_id'],'quote':x['text']} for x in c.get('content_candidates',[])],
+            'source_urls':[s['source_url'] for s in c['sources']]}
+
+
+def complete(answer):
+    return lambda _: {'choices':[{'message':{'content':json.dumps(answer)}}]}
+
+
+@pytest.mark.parametrize('question,facets',[
+    ('Aké problémy riešili Eva Nová a Boris Nový a aké experimenty mali pripraviť?',{'problems','experiments'}),
+    ('Aké metódy používa Eva Nová?',{'methods'}),
+    ('Aké sú dosiahnuté výsledky Evy Novej?',{'results'}),
+])
+def test_content_without_explicit_work_type(corpus,question,facets):
+    payload=prepare(corpus,question); c=rag_evidence_context(question,payload)
+    assert c['requirements']['content'] and set(c['requirements']['facets'])==facets
+    assert c['content_candidates']
+
+
+def test_multiple_people_and_aspects_remain_bound(corpus):
+    q='Aké problémy riešili Eva Nová a Boris Nový a aké experimenty mali pripraviť?'
+    payload=prepare(corpus,q); a=model_selection(q,payload)
+    r=generate_verified_rag(q,payload,complete(a))
+    assert r['accepted_for_delivery'] and r['status']=='ok'
+    text=r['answer']; eva,boris=text.split('Boris Nový — bakalárska práca')
+    assert 'tokenizérom' in eva and 'PDF' not in eva
+    assert 'PDF' in boris and 'tokenizérmi' not in boris
+    assert text.count('**Riešené problémy:**')==2
+    assert text.count('**Plánované experimenty a vyhodnotenie:**')==2
+
+
+def test_ambiguous_work_is_a_clarification_without_model(corpus):
+    db,raw=corpus; source=raw['sources'][0]
+    with sqlite3.connect(db) as conn:
+        conn.execute('INSERT INTO chunks VALUES(?,?,?,?,?)',(source['document_path'],
+            'Stav:\nVznikol iný problém.',json.dumps([['Eva Nová','Bakalárska práca 2022']]),1,1))
+    q='Aké problémy mala Eva Nová počas práce?'
+    p=prepare(corpus,q)
+    r=generate_verified_rag(q,p,lambda _:pytest.fail('No model call for a clarification'))
+    assert r['status']=='ambiguous' and r['reason']=='ambiguous_work'
+    assert '2022' in r['answer'] and '2024' in r['answer']
+
+
+def test_explicit_type_and_year_resolve_scope(corpus):
+    db,raw=corpus
+    with sqlite3.connect(db) as conn:
+        conn.execute('INSERT INTO chunks VALUES(?,?,?,?,?)',(raw['sources'][0]['document_path'],
+            'Stav:\nChyba v inom projekte Cassandra.',json.dumps([['Eva Nová','Diplomový projekt 2022']]),1,1))
+    q='Aké problémy mala Eva Nová pri diplomovej práci v roku 2024?'
+    p=prepare(corpus,q)
+    assert p['status']=='ok' and 'Cassandra' not in p['context']
+
+
+def test_year_condition_without_the_words_v_roku_is_preserved(corpus):
+    q='Aké problémy mala Eva Nová pri diplomovej práci 1999?'
+    p=prepare(corpus,q); c=rag_evidence_context(q,p)
+    assert c['requirements']['years']==[1999] and not c.get('content_candidates')
+    r=generate_verified_rag(q,p,lambda _:pytest.fail('No model for missing scoped evidence'))
+    assert r['accepted_for_delivery'] and r['status']=='no_answer' and 'tokenizérom' not in r['answer']
+    assert requested_years('Práca „Model 2020“ z roku 2024')==[2024]
+
+
+def test_named_source_can_be_loaded_when_absent_from_top_results(corpus):
+    db,raw=corpus; raw['sources']=raw['sources'][:1];raw['context']=build_context_text(raw['sources'])
+    p=prepare(corpus,'Aké problémy riešili Eva Nová a Boris Nový?')
+    assert {s['title'] for s in p['sources']}=={'Eva Nová','Boris Nový'}
+
+
+def test_preparation_does_not_change_the_original_retrieval(corpus):
+    db,raw=corpus; before=copy.deepcopy(raw)
+    prepare_named_content(db,'Aké problémy riešili Eva Nová a Boris Nový?',raw)
+    assert raw==before
+
+
+def test_source_budget_cannot_silently_remove_a_requested_person(corpus):
+    db,raw=corpus; q='Aké problémy riešili Eva Nová a Boris Nový?'
+    p=prepare_named_content(db,q,raw)
+    p['sources'][0]['text']+='\n'+'Veľmi dlhý zdrojový text. '*1000
+    compact=compact_rag_payload(p,max_tokens=1100)
+    c=rag_evidence_context(q,compact)
+    assert compact['truncated']
+    assert set(c['requirements']['people'])=={'Eva Nová','Boris Nový'}
+
+
+def test_candidates_are_actual_bound_literal_evidence(corpus):
+    q='Aké problémy riešili Eva Nová a Boris Nový?'; c=rag_evidence_context(q,prepare(corpus,q))
+    for candidate in candidates(c):
+        assert any(all(candidate[k]==p[k] for k in ['person','work_type','year','source_id'])
+                   and candidate['text'] in p['text'] for p in c['content_passages'])
+
+
+def test_unknown_second_subject_is_not_silently_ignored(corpus):
+    p=prepare(corpus,'Aké problémy riešili študenti Eva Nová a Boris Noyy?')
+    assert p['status']=='ambiguous' and p['reason']=='unresolved_subject'
+
+
+@pytest.mark.parametrize('question',[
+    'Aké problémy riešili Eva Nová a Xaver Neexistujúci?',
+    'Aké problémy riešil Boris Noyy?',
+])
+def test_unresolved_subject_cannot_be_replaced_by_available_profiles(corpus,question):
+    p=prepare(corpus,question)
+    assert p['status']=='ambiguous' and p['reason']=='unresolved_subject'
+
+
+def test_partial_answer_names_the_missing_aspect(corpus):
+    db,raw=corpus
+    with sqlite3.connect(db) as conn:
+        conn.execute("UPDATE chunks SET text='Stav:\nVznikol problém s parserom.' WHERE document_path LIKE '%boris_novy%'")
+    q='Aké problémy riešili Eva Nová a Boris Nový a aké experimenty mali pripraviť?'
+    p=prepare(corpus,q); r=generate_verified_rag(q,p,complete(model_selection(q,p)))
+    assert r['accepted_for_delivery'] and r['answer_coverage']['status']=='partial'
+    assert 'Táto časť odpovede nie je v poskytnutých podkladoch doložená.' in r['answer']
+    assert 'parserom' in r['answer']
+
+
+def test_all_aspects_absent_is_no_answer_not_rejection(corpus):
+    db,raw=corpus
+    with sqlite3.connect(db) as conn:
+        conn.execute("UPDATE chunks SET text='Zadanie:\n- Implementujte parser.'")
+    q='Aké problémy riešili Eva Nová a Boris Nový?'
+    r=generate_verified_rag(q,prepare(corpus,q),lambda _:pytest.fail('No model for a bounded empty result'))
+    assert r['status']=='no_answer' and r['accepted_for_delivery'] and not r['claims']
+
+
+def test_known_year_is_returned_when_requested_title_is_missing(corpus):
+    _,raw=corpus; source=raw['sources'][0]
+    source['text']='Sekcia: Diplomová práca 2024\nStav:\nImplementovaný parser.'
+    raw['sources']=[source];raw['context']=build_context_text(raw['sources'])
+    q='Uveď názov a rok diplomovej práce Evy Novej.'
+    a={'status':'ok','answer':'','claims':[{'person':'Eva Nová','work_type':'diploma_thesis','year':2024}],
+       'evidence_quotes':[],'source_urls':[source['source_url']]}
+    r=generate_verified_rag(q,raw,complete(a))
+    assert r['accepted_for_delivery'] and '2024' in r['answer']
+    assert 'názov práce nie je v poskytnutých podkladoch doložený' in r['answer']
+    assert r['answer_coverage']['status']=='partial' and not r['claims'][0].get('title')
+
+
+def test_concrete_results_win_over_administrative_notes(corpus):
+    q='Aké výsledky dosiahla Eva Nová?'; p=prepare(corpus,q); c=rag_evidence_context(q,p)
+    text='\n'.join(x['text'] for x in c['content_candidates'])
+    assert 'Natrénovaný model BERT' in text
+    r=generate_verified_rag(q,p,complete(model_selection(q,p)))
+    assert r['accepted_for_delivery'] and 'Natrénovaný model BERT' in r['answer']
+    assert 'Konzultácia' not in r['answer']
+
+
+def test_duplicate_quotes_and_overlapping_chunks_do_not_duplicate_output(corpus):
+    q='Aké výsledky dosiahla Eva Nová?'; p=prepare(corpus,q); a=model_selection(q,p)
+    a['evidence_quotes']*=2
+    r=generate_verified_rag(q,p,complete(a))
+    assert r['accepted_for_delivery'] and r['answer'].count('Natrénovaný model BERT')==1
+    assert 'Ďalšie podklady:' not in r['answer'] and 'Ďalšie doložené podklady' not in r['answer']
+
+
+def test_plans_cannot_answer_actual_use_or_completed_experiments():
+    assert 'methods' not in facets_for('Použite metódu BERT.','task','Aké metódy skutočne používal?')
+    assert 'experiments' not in facets_for('Pripravte experiment.','task','Aké experimenty vykonal?')
+    assert 'results' not in facets_for('Cieľom je implementovať systém.','progress','Aké dosiahol výsledky?')
+
+
+def test_method_answer_does_not_add_valid_but_unrequested_profile_metadata(corpus):
+    _,raw=corpus
+    raw['sources'][0]['author']='Daniel Hladek'
+    q='Aké metódy používala Eva Nová pri diplomovej práci v roku 2024?'
+    p=prepare(corpus,q); a=model_selection(q,p)
+    a['claims']=[{'person':'Eva Nová','work_type':'diploma_thesis','year':2024,
+                  'document_author':'Daniel Hladek'}]
+    r=generate_verified_rag(q,p,complete(a))
+    assert r['accepted_for_delivery'] and 'BERT' in r['answer']
+    assert 'Daniel Hladek' not in r['answer'] and 'autor dokumentu' not in r['answer']
+    assert r['answer'].count('2024')==1
+
+
+@pytest.mark.parametrize('question',[
+    'Práce od roku 2020 do roku 2024?', 'Práce pred rokom 2020?',
+    'Študenti s aspoň 3 prácami?', 'Práce okrem diplomových?', 'Študenti bez tagu RAG?',
+    'V ktorom roku Ján Holp začal pracovať na diplomovej práci?',
+    'Kedy Ján Holp obhájil diplomovú prácu?',
+])
+def test_unsupported_operators_are_not_ignored(question):
+    assert unsupported_constraint(question)
+    result=gs.search_graph_with_session(FakeSession({}),question)
+    assert result['status']=='ambiguous' and result['reason']=='unsupported_constraint' and not result['facts']
+
+
+@pytest.mark.parametrize('question',[
+    'Práca v roku 1999?', 'Aký je názov práce „Od 2020 do 2024“?',
+    'Aké práce alebo projekty má Eva Nová?',
+])
+def test_supported_year_and_work_alternative_are_not_blocked(question):
+    assert unsupported_constraint(question) is None
+
+
+@pytest.mark.parametrize('reason',['model_unavailable','retrieval_unavailable','answer_rejected','answer_capacity_exceeded'])
+def test_public_errors_are_distinct_and_do_not_leak_rejected_text(reason):
+    r=public_failure({'reason':reason,'answer':'SECRET YEAR 2099','claims':[{'person':'Eva'}],'source_urls':['SECRET']})
+    assert r['reason']==reason and not r['accepted_for_delivery']
+    assert 'SECRET' not in r['answer'] and '2099' not in r['answer']
+    assert not r['claims'] and not r['source_urls']
+
+
+def test_retrieval_failure_is_not_misreported_as_bad_model_answer(monkeypatch):
+    import tools.answers.answer_delivery as delivery
+    def unavailable(_): raise RuntimeError('SECRET backend address')
+    monkeypatch.setattr(delivery,'search_graph',unavailable)
+    r=delivery.deliver_answer('Aké práce má Eva Nová?',engine='graphrag',
+                             complete=lambda _:pytest.fail('No model call after failed retrieval'))
+    assert r['status']=='error' and r['reason']=='retrieval_unavailable'
+    assert 'SECRET' not in r['answer']

tools/answer_evaluation/answer_coverage.py

@@ -39,7 +39,7 @@ def answer_coverage(answer: dict, context: dict) -> dict:
                      (row.get('source_url') in cited or answer.get('status') == 'no_answer')))
     # A context budget may drop only unrelated sources. A clipping marker in
     # the actual cited block is stronger evidence than this global reason.
-    context_limited = 'llm_context_budget' in reasons
+    context_limited = any(reason in reasons for reason in ['llm_context_budget','answer_candidate_budget','evidence_unit_limit'])
     limited = context.get('complete') is not True or context.get('truncated') is True
     limitations = [*reasons]
     if clipped:
@@ -48,6 +48,13 @@ def answer_coverage(answer: dict, context: dict) -> dict:
     status = answer.get('status')
     if status == 'ok':
         coverage = 'supported'
+        if any(requirements.get(key) and key not in fields for fields in requirements.get('available_fields',{}).values()
+               for key in ['title','year','start_year']):
+            coverage='partial'; warnings.append('partial_requested_fields')
+        if requirements.get('facets'):
+            from .content_selection import missing_facets
+            if missing_facets(answer,context):
+                coverage='partial'; warnings.append('partial_requested_aspects')
         if scope == 'list' and (limited or clipped):
             coverage = 'partial'; warnings.append('list_may_be_incomplete')
         elif scope == 'count' and (context.get('aggregation') or {}).get('complete') is not True:
@@ -74,4 +81,7 @@ def coverage_notices(coverage: dict) -> list[str]:
         'missing_information_with_clipped_context': 'Časť podkladov bola pri príprave kontextu skrátená.',
         'count_not_established': 'Celkový počet sa z dostupných podkladov nedá spoľahlivo určiť.',
     }
-    return [messages[reason] for reason in coverage['warnings']]
+    notices=[messages[reason] for reason in coverage['warnings'] if reason in messages]
+    if 'list_may_be_incomplete' in coverage['warnings']:
+        notices.append('Výpis môžeš zúžiť uvedením osoby, typu práce alebo konkrétneho roku.')
+    return notices

tools/answer_evaluation/checked_generation.py

@@ -16,7 +16,8 @@ def generate_checked_answer(question: str, context: dict, prompt: str, complete)
             result['response']=response;record['response']=response
             text=response['choices'][0]['message']['content'].strip()
         except Exception as error:
-            result['error']=f'{type(error).__name__}: {str(error)[:500]}'
+            result['error']=type(error).__name__
+            result['reason']='model_unavailable'
             result['attempts'].append(record)
             return result
         result['attempts'].append(record)

tools/answer_evaluation/content_selection.py

@@ -0,0 +1,191 @@
+"""Role-aware, bounded excerpt selection and presentation from actual evidence."""
+from __future__ import annotations
+
+import re
+from .context_policy import normalized
+from .question_intent import FACET_LABELS, planned_experiments
+
+COMPLETED = r'\b(?:implementovan\w*|natrenovan\w*|vytvoren\w*|extrahovan\w*|vykonan\w*|urobene vyhodnotenie|vyhodnoteny\w*|nameran\w*)\b'
+PLANNED = r'\b(?:bude|budu|planuj\w*|cielom|mal by|ma sa|treba|pripravte|navrhnite|otestujte|vyhodnotte)\b'
+ADMIN = r'\b(?:konzultac\w*|struktura prace|pisomn\w*|pisat|odovzda\w*|wip|termin\w*)\b'
+
+
+def split_units(text: str) -> list[str]:
+    from .rag_evidence import descriptive_text
+    parts = re.split(r'\n\s*\n|(?=^[ \t]*(?:\d+[.)]|[-+*])\s+)', text, flags=re.M)
+    units = []
+    for part in parts:
+        part = part.strip()
+        if not descriptive_text(part):
+            continue
+        # A source paragraph is kept literal, including its negations and qualifications.
+        for unit in ([part] if len(part) <= 900 else re.split(r'(?<=[.!?])\s+(?=[A-ZÁČĎÉÍĽŇÓŠŤÚÝŽ])', part)):
+            if descriptive_text(unit):
+                units.append(unit.strip())
+    return units
+
+
+def effective_role(text: str, role: str) -> str:
+    n = normalized(text)
+    if role=='other':
+        from .presentation import GOAL_VERBS
+        first=n.split()[0] if n else ''
+        if first in {normalized(word) for word in GOAL_VERBS} or n.startswith('cielom je'):
+            return 'task'
+    if role == 'other' and re.search(COMPLETED, n) and not re.search(PLANNED, n):
+        return 'progress'
+    return role
+
+
+def facets_for(text: str, role: str, question: str) -> list[str]:
+    n = normalized(text); facets = []
+    if role == 'assignment': facets.append('assignment')
+    if role == 'progress' and (not re.search(PLANNED,n) or re.search(COMPLETED,n)):
+        facets.append('results')
+    if re.search(r'\b(?:problem\w*|tazkost\w*|prekazk\w*|chyb\w*|zlyha\w*|nepodar\w*|nevyries\w*)\b', n) and role not in {'assignment','task'}:
+        facets.append('problems')
+    if re.search(r'\b(?:metod\w*|technolog\w*|nastroj\w*|pouzi\w*|vyuziv\w*|pomocou|implementovan\w*|natrenovan\w* model\w*)\b', n):
+        if role not in {'assignment','task'} or not re.search(r'\b(?:skutocne|pouziva\w*|pouzil\w*|vyuzival\w*|implementoval\w*)\b',normalized(question)):
+            facets.append('methods')
+    if re.search(r'\b(?:experiment\w*|meran\w*|test\w*|otest\w*|vyhodnot\w*|benchmark\w*|f1|presnost\w*)\b', n):
+        wants_done=re.search(r'\b(?:vykonal\w*|otestoval\w*|uskutocnil\w*|vyhodnotil\w*)\b',normalized(question))
+        if wants_done and role in {'assignment','task'}:
+            pass
+        elif not planned_experiments(question) or role in {'assignment','task'} or re.search(PLANNED, n):
+            facets.append('experiments')
+    return facets
+
+
+def unit_score(text: str, role: str, facet: str = '') -> int:
+    n = normalized(text)
+    score = 25 if role == 'progress' else 10
+    if re.search(COMPLETED, n): score += 45
+    if re.search(r'\b(?:model|system|algoritm\w*|vyhodnot\w*|experiment\w*|presnost\w*|pagerank|bert)\b', n): score += 15
+    if re.search(ADMIN, n): score -= 70
+    if facet == 'problems' and re.search(r'\b(?:problem\w*|chyb\w*|zlyha\w*)\b', n): score += 35
+    if role in {'assignment','task'} and facet == 'results': score -= 100
+    return score
+
+
+def evidence_units(context: dict, quotes: list[dict] | None = None) -> list[dict]:
+    from .rag_evidence import content_sections, quote_sections
+    rows = []
+    for passage in context.get('content_passages', []):
+        pieces = content_sections(passage['text']) if quotes is None else [
+            piece for quote in quotes if quote['source_id'] == passage['source_id'] and
+            ' '.join(quote['quote'].split()) in ' '.join(passage['text'].split())
+            for piece in quote_sections(quote, context)]
+        for piece in pieces:
+            for unit in split_units(piece['text']):
+                role = effective_role(unit, piece['kind'])
+                rows.append({k: passage.get(k) for k in ['person','work_type','year','source_id']} |
+                            {'text':unit,'kind':role,'facets':facets_for(unit,role,context.get('question',''))})
+    # Overlapping chunks/quotes can repeat the same fact with different roles.
+    # An explicit role wins over an unlabelled duplicate, within the same work.
+    unique = {}
+    for row in rows:
+        clean=re.sub(r'^\s*(?:\d+[.)]|[-+*])\s+','',row['text'])
+        key = tuple(row.get(k) for k in ['person','work_type','year']) + (' '.join(clean.split()),)
+        old = unique.get(key)
+        if old is None or (old['kind']=='other' and row['kind']!='other'):
+            unique[key]=row
+    return list(unique.values())
+
+
+def candidates(context: dict) -> list[dict]:
+    facets = context['requirements'].get('facets', [])
+    selected = []; seen = set(); units = evidence_units(context)
+    for person in context['requirements']['people']:
+        for facet in facets:
+            relevant = [u for u in units if u['person']==person and facet in u['facets']]
+            relevant.sort(key=lambda u: -unit_score(u['text'],u['kind'],facet))
+            accepted=0
+            for unit in relevant:
+                text = unit['text']
+                if len(text)>2500:
+                    context['complete']=False; context['truncated']=True
+                    context['truncation_reasons']=list(dict.fromkeys([*(context.get('truncation_reasons') or []),'evidence_unit_limit']))
+                    continue
+                key=(unit['source_id'],text)
+                if key not in seen:
+                    seen.add(key); selected.append({**unit,'text':text})
+                accepted+=1
+                if accepted==2: break
+    return selected
+
+
+def missing_facets(answer: dict, context: dict) -> list[tuple[str,str]]:
+    found = evidence_units(context, answer.get('evidence_quotes', []))
+    return [(person,facet) for person in context['requirements']['people']
+            for facet in context['requirements'].get('facets', [])
+            if not any(row['person']==person and facet in row['facets'] for row in found)]
+
+
+def fit_candidates(sent: dict, context: dict, limit: int, encoding: str) -> None:
+    """Budget the actual candidate JSON, sharing space across named people."""
+    import json
+    from tools.rag.build_chunks import TokenCounter
+    counter=TokenCounter(encoding)
+    if counter.encoding is None:
+        raise RuntimeError('Missing context tokenizer')
+    original=list(context.get('content_candidates',[]))
+    groups=[[row for row in original if row.get('person')==person]
+            for person in context['requirements']['people']]
+    ordered=[group[i] for i in range(max(map(len,groups),default=0)) for group in groups if i<len(group)]
+    selected=[]
+    for row in ordered:
+        if counter.count(json.dumps({**sent,'content_candidates':[*selected,row]},ensure_ascii=False))<=limit:
+            selected.append(row)
+    if len(selected)!=len(original):
+        context['complete']=False;context['truncated']=True
+        context['truncation_reasons']=list(dict.fromkeys([*(context.get('truncation_reasons') or []),'answer_candidate_budget']))
+        sent.update({key:context[key] for key in ['complete','truncated','truncation_reasons']})
+    context['content_candidates']=selected;sent['content_candidates']=selected
+    # Account for the added limitation metadata as well.
+    while selected and counter.count(json.dumps(sent,ensure_ascii=False))>limit:
+        selected.pop()
+    if counter.count(json.dumps(sent,ensure_ascii=False))>limit:
+        raise ValueError('Mandatory answer metadata exceeds context budget')
+
+
+def facet_lines(answer: dict, context: dict) -> list[str]:
+    from .presentation import goal_text, work_subject, owner_in_question
+    from .rag_evidence import WORK_LABELS
+    req=context['requirements']; units=evidence_units(context,answer.get('evidence_quotes',[]))
+    lines=[]
+    for person in req['people']:
+        scopes={(p['work_type'],p.get('year')) for p in context.get('content_passages',[]) if p.get('person')==person}
+        kind=req['work_types'].get(person) or (next(iter(scopes))[0] if len(scopes)==1 else None)
+        subject=work_subject(person,kind,context.get('question',''))
+        if not owner_in_question(person,context.get('question','')):
+            subject=person+(' — '+WORK_LABELS[kind] if kind in WORK_LABELS else '')
+        if len(scopes)==1 and next(iter(scopes))[1] is not None:
+            subject+=' ('+str(next(iter(scopes))[1])+')'
+        lines.extend(['**'+subject+'**', ''])
+        seen=set()
+        for facet in req.get('facets', []):
+            label=FACET_LABELS[facet]
+            if facet=='experiments' and planned_experiments(context.get('question','')):
+                label='Plánované experimenty a vyhodnotenie'
+            rows=[u for u in units if u['person']==person and facet in u['facets']]
+            rows.sort(key=lambda u:-unit_score(u['text'],u['kind'],facet))
+            lines.append('**'+label+':**')
+            shown=0
+            for row in rows:
+                key=' '.join(re.sub(r'^\s*(?:\d+[.)]|[-+*])\s+','',row['text']).split())
+                if key in seen:
+                    continue
+                text=row['text']
+                if row['kind'] in {'assignment','task'}:
+                    text,_=goal_text(text)
+                text=re.sub(r'^\s*(?:\d+[.)]|[-+*])\s+','',text)
+                phase=' (plán)' if row['kind'] in {'assignment','task'} and facet not in {'assignment'} else ''
+                lines.append('- '+text+phase); seen.add(key); shown+=1
+                if shown==2: break
+            if not shown:
+                if rows:
+                    lines.append('Je doložené rovnakým výňatkom uvedeným vyššie.')
+                else:
+                    lines.append('Táto časť odpovede nie je v poskytnutých podkladoch doložená.')
+            lines.append('')
+    return lines

tools/answer_evaluation/graph_context.py

@@ -54,6 +54,7 @@ def graph_context(result: dict) -> dict:
              'index_version':result.get('index_version')}
     context['answer_candidates']=joined_candidates(facts,intent)
     context['question'] = result.get('query', '')
+    context['clarification'] = result.get('clarification')
     context['requested_people'] = list(dict.fromkeys(row['name']
         for row in result.get('resolved_entities', []) if row.get('kind') == 'Person' and row.get('name')))
     comparison = result.get('academic_year_comparison') or next((row for row in result['facts']

tools/answer_evaluation/question_intent.py

@@ -0,0 +1,38 @@
+"""Explicit answer aspects; these select evidence, never supply facts."""
+import re
+from .context_policy import normalized
+
+FACET_PATTERNS = {
+    'assignment': r'\b(?:zadani\w*|ciel\w*)\b',
+    'problems': r'\b(?:problem\w*|tazkost\w*|prekazk\w*|chyb\w*)\b',
+    'methods': r'\b(?:metod\w*|technolog\w*|nastroj\w*|pouziv\w*|pouzil\w*)\b',
+    'experiments': r'\b(?:experiment\w*|merani\w*|testovani\w*|vyhodnocov\w*)\b',
+    'results': r'\b(?:vysledk\w*|dosiahn\w*|stav\w*)\b',
+}
+FACET_LABELS = {'assignment': 'Zadanie a ciele', 'problems': 'Riešené problémy',
+                'methods': 'Metódy a technológie', 'experiments': 'Experimenty a vyhodnotenie',
+                'results': 'Zaznamenaný stav a výsledky'}
+
+
+def requested_facets(question: str) -> list[str]:
+    text = normalized(re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', question))
+    matches = [(m.start(), key) for key, pattern in FACET_PATTERNS.items()
+               if (m := re.search(pattern, text))]
+    return [key for _, key in sorted(matches)]
+
+
+def content_question(question: str) -> bool:
+    text = normalized(question)
+    return bool(requested_facets(question) or re.search(
+        r'\b(?:comu|o com|obsah\w*|opis\w*|popis\w*|zamer\w*|zhrn\w*|implement\w*)\b', text))
+
+
+def facet_mode(question: str) -> bool:
+    facets = requested_facets(question)
+    return any(facet in facets for facet in ['problems','methods','experiments','results'])
+
+
+def planned_experiments(question: str) -> bool:
+    text=normalized(question)
+    return bool(re.search(r'\b(?:plan\w*|pripravit|navrhnut)|'
+                          r'\b(?:mal|mali|mala|ma|maju)\s+(?:priprav\w*|navrhn\w*|otest\w*|vyhodnot\w*)',text))

tools/answer_evaluation/rag_evidence.py

@@ -10,6 +10,8 @@ import json
 import sqlite3
 from contextlib import closing
 from .context_policy import normalized, text_contains_value, person_pattern
+from .question_intent import requested_facets, content_question, facet_mode
+from tools.retrieval.question_constraints import requested_years
 
 WORK_TYPES = {
     'bakalarska praca': 'bachelor_thesis', 'diplomova praca': 'diploma_thesis',
@@ -148,6 +150,10 @@ def academic_windows(body: str) -> list[tuple[tuple[str, int | None] | None, str
                 scope = (None if normalized(heading[1]).startswith(('priprava na ', 'priprava k ')) else
                          detected or (heading_stack[-1][1] if heading_stack else None))
                 heading_stack.append((level, scope))
+                if scope is not None and detected is None:
+                    # Keep semantic child headings (Stav, Úlohy, Metódy) as
+                    # literal source text instead of losing their meaning.
+                    lines.append(line)
         else:
             lines.append(line)
     flush()
@@ -392,14 +398,16 @@ def rag_evidence_context(question: str, payload: dict) -> dict:
             facts.extend({**base, field: value} for value in values)
     facts = list({tuple(sorted(row.items())): row for row in facts}.values())
     requested = [person for person in people if text_contains_value(question, person, person=True)]
+    requested=list(dict.fromkeys([*payload.get('named_subjects',[]),*requested]))
     expected_types = requested_types(question, requested)
     query = normalized(question)
     intent_query = normalized(re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', question))
     academic = bool(TYPE_PATTERN.search(query) or re.search(r'\b(?:rok\w* prac\w*|nazov\w* prac\w*)\b', query))
-    content = bool(academic or (requested and re.search(r'\b(?:prac\w*|projekt\w*)\b', query))) and bool(re.search(
+    content = bool(academic or requested) and bool(re.search(
         r'\b(?:comu(?:\s+sa)?\s+ven(?:uj|ov)\w*|o\s+com|co\s+ries\w*|obsah\w*|opis\w*|popis\w*|'
         r'zhrn\w*|zhrnut\w*|vysvetl\w*|zamer\w*|zadani\w*|ciel\w*|metod\w*|technolog\w*|'
-        r'pouziv\w*|implement\w*|vysledk\w*|dosiahn\w*)\b', intent_query))
+        r'pouziv\w*|implement\w*|vysledk\w*|dosiahn\w*|problem\w*|tazkost\w*|prekazk\w*|experiment\w*|merani\w*)\b', intent_query)
+        or content_question(question))
     identity_first = re.search(r'^(?:aky|aka|ake|ktory|ktora|ktore|v ktorom|uved)'
         r'(?:\s+(?:je|su|bol|bola|boli|presny|presne)){0,3}\s+(?:nazov\w*|rok\w*)\b', intent_query)
     explicit_description = re.search(r'\b(?:comu|o com|opis\w*|popis\w*|zhrn\w*|obsah\w*|zadani\w*|ciel\w*)\b', intent_query)
@@ -419,18 +427,27 @@ def rag_evidence_context(question: str, payload: dict) -> dict:
         requirements['response_mode'] = 'mixed' if any(requirements[key] for key in ['title', 'year', 'start_year']) else 'quotes'
         requirements['distinguish_progress'] = bool(re.search(r'\b(?:zadani\w*|ciel\w*)\b', query)
             and re.search(r'\b(?:vysledk\w*|dosiahn\w*|stav\w*)\b', query))
+        if facet_mode(question):
+            requirements['facets'] = requested_facets(question)
     stated_years = re.findall(r'\b(?:v roku|z roku|pre rok|na rok|roku|rok)\s+(\d+)\b', query) if academic else []
-    requirements['years'] = ([int(value) for value in stated_years if len(value) == 4]
-                              if not requirements['start_year'] else [])
+    requirements['years'] = requested_years(question) if academic and not requirements['start_year'] else []
     requirements['unsupported_year'] = any(len(value) != 4 for value in stated_years)
     if requested:
         facts = [row for row in facts if row['person'] in requested]
-    context = {'question': question, **{key: payload.get(key) for key in ['status', 'reason', 'complete', 'truncated',
+    requirements['available_fields'] = {person:[key for key in ['title','year','start_year']
+        if any(row.get('person')==person and row.get(key) is not None and
+               (key=='start_year' or ((not expected_types.get(person) or row.get('work_type')==expected_types[person])
+                and (not requirements['years'] or row.get('year') in requirements['years']))) for row in facts)]
+        for person in requested}
+    context = {'question': question, 'clarification':payload.get('clarification'), **{key: payload.get(key) for key in ['status', 'reason', 'complete', 'truncated',
              'truncation_reasons', 'index_version']}, 'sources': payload.get('sources', []),
              'facts': facts, 'source_blocks': blocks, 'requirements': requirements,
              'aggregation': None, 'answer_candidates': []}
     if content:
         context['content_passages'] = academic_content_passages(blocks, facts, requirements, question)
+        if requirements.get('facets'):
+            from .content_selection import candidates
+            context['content_candidates'] = candidates(context)
         if requirements['distinguish_progress']:
             requirements['content_roles'] = {person: sorted({section['kind']
                 for p in context['content_passages'] if p['person'] == person

tools/answer_evaluation/rag_preparation.py

@@ -0,0 +1,91 @@
+"""Resolve named source profiles and scope content before the shared token budget."""
+from __future__ import annotations
+
+import json
+import re
+import sqlite3
+from contextlib import closing
+
+from .context_policy import normalized, person_pattern, text_contains_value
+from .question_intent import content_question
+from .rag_evidence import PROFILE, academic_scope, academic_windows, requested_types, WORK_LABELS
+from tools.retrieval.question_constraints import requested_years
+
+
+def clarify(payload: dict, reason: str, message: str) -> dict:
+    return {**payload,'status':'ambiguous','reason':reason,'complete':False,
+            'clarification':message}
+
+
+def prepare_named_content(db_file, question: str, payload: dict) -> dict:
+    if not content_question(question):
+        return payload
+    if re.search(r'\b(?:s tagom|s klucovym slovom|z kategorie|s autorom dokumentu)\b',normalized(question)):
+        return clarify(payload,'content_filter_requires_graph',
+            'Najprv over podmienku tagu, kategórie alebo autora cez /graphrag. '
+            'Potom môžem opísať obsah konkrétnej vyhovujúcej práce; túto podmienku nevynechám.')
+    from tools.rag.rag_utils import build_context_text
+    with closing(sqlite3.connect(db_file.resolve().as_uri()+'?mode=ro',uri=True)) as conn:
+        conn.row_factory=sqlite3.Row
+        if not {'documents','chunks'} <= {r[0] for r in conn.execute("SELECT name FROM sqlite_master WHERE type='table'")}:
+            return payload
+        catalog=[dict(r) for r in conn.execute('SELECT path,title,author,published FROM documents WHERE COALESCE(published,1)=1')
+                 if PROFILE.fullmatch(r['path']) and r['title']]
+        named=[r for r in catalog if text_contains_value(question,r['title'],person=True)]
+        # Preserve question order rather than the original retrieval ranking.
+        named.sort(key=lambda r:re.search(person_pattern(r['title']),normalized(question)).start())
+        if len(named)>4:
+            return clarify(payload,'too_many_subjects','Pre podrobné obsahové porovnanie vyber najviac štyri osoby naraz.')
+        from tools.graphrag.graph_search import requested_name_phrases
+        first_names={normalized(r['title']).split()[0] for r in catalog}
+        capitalized={normalized(match[0]) for match in re.finditer(
+            r'\b[A-ZÁÄČĎÉÍĹĽŇÓÔŔŠŤÚÝŽ][^\W\d_]+(?:\s+[A-ZÁÄČĎÉÍĹĽŇÓÔŔŠŤÚÝŽ][^\W\d_]+){1,3}\b',question)}
+        unknown=[]
+        for phrase in requested_name_phrases(question):
+            words=phrase.split()
+            named_companion=(named and phrase in capitalized and
+                             re.search(r'\ba\s+'+re.escape(phrase)+r'\b',normalized(question)))
+            if (2<=len(words)<=4 and (words[0] in first_names or named_companion) and
+                    not any(text_contains_value(phrase,r['title'],person=True) for r in catalog)):
+                unknown.append(phrase)
+        if unknown:
+            return clarify(payload,'unresolved_subject','Meno „'+unknown[0]+'“ sa nepodarilo spoľahlivo priradiť. '
+                           'Uveď jeho presné celé meno; ostatné osoby nebudem používať ako náhradu.')
+        if not named:
+            return payload
+        types=requested_types(question,[r['title'] for r in named])
+        years=requested_years(question)
+        selected=[]
+        for person in named:
+            rows=list(conn.execute('SELECT text,heading_paths_json FROM chunks WHERE document_path=? '
+                                   'AND COALESCE(published,1)=1 ORDER BY chunk_index',(person['path'],)))
+            frames=[]; records={};seen=set()
+            for row in rows:
+                paths=json.loads(row['heading_paths_json'] or '[]')
+                headings=[' > '.join(p) if isinstance(p,list) else str(p) for p in paths]
+                first=headings[0] if headings else ''
+                for scope,text in academic_windows('Sekcia: '+first+'\n'+str(row['text'] or '')):
+                    if not scope or (types.get(person['title']) and scope[0]!=types[person['title']]) or (years and scope[1] not in years):
+                        continue
+                    label=WORK_LABELS[scope[0]]+(f' {scope[1]}' if scope[1] is not None else '')
+                    records[scope]=label
+                    if (scope,text) not in seen:
+                        seen.add((scope,text));frames.append('Sekcia: '+label+'\n'+text)
+            if len(records)>1:
+                options='; '.join(dict.fromkeys(records.values()))
+                return clarify(payload,'ambiguous_work',f'Ktorú prácu osoby {person["title"]} máš na mysli? '
+                               f'V indexe sú tieto možnosti: {options}. Uveď typ práce a rok.')
+            source=next((s for s in payload.get('sources',[]) if s.get('document_path')==person['path']),None)
+            if source is None:
+                from tools.rag.search_core import make_source_url
+                source={'title':person['title'],'document_path':person['path'],'source_url':make_source_url(person['path']),
+                        'author':person['author'],'published':person['published']}
+            selected.append({**source,'text':'\n\n'.join(frames)})
+        # Keep already-retrieved child documents of the requested profiles.
+        for source in payload.get('sources',[]):
+            if any(source.get('document_path','').startswith(p['path'].removesuffix('README.md'))
+                   and source.get('document_path')!=p['path'] for p in named):
+                selected.append(source)
+        selected=[{**source,'source_id':f'S{i}'} for i,source in enumerate(selected,1)]
+        return {**payload,'sources':selected,'source_count':len(selected),'context':build_context_text(selected),
+                'named_subjects':[p['title'] for p in named]}

tools/answer_evaluation/render_answer.py

@@ -54,6 +54,8 @@ def render_graph_answer(answer: dict, context: dict) -> dict:
             lines.append(f'Počet vyhovujúcich osôb: {out["count"]}.')
         if context.get('complete') is False or context.get('truncated'):
             lines.append('Výpis je neúplný.')
+            if records:
+                lines.append('Výpis môžeš zúžiť uvedením osoby, typu práce alebo konkrétneho roku.')
         elif context.get('complete') is None:
             lines.append('Úplnosť vyhľadania nemožno zaručiť.')
     elif out['status'] == 'ambiguous':

tools/answer_evaluation/verified_rag.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import json
+import os
 import re
 from typing import Literal
 
@@ -12,6 +13,7 @@ from .context_policy import answer_context_issues, citation_check, allowed_urls,
 from .rag_evidence import descriptive_text, rag_evidence_context, quote_sections
 from .answer_coverage import answer_coverage, coverage_notices
 from .presentation import content_lines, fact_sentences, no_answer_text, work_subject
+from .content_selection import facet_lines, evidence_units, missing_facets, fit_candidates
 
 
 class RagClaim(Claim):
@@ -46,6 +48,8 @@ uveď start_year aj year a work_type. Citácia ani nesúvisiaci výňatok údaj
 Rešpektuj requirements.response_mode. Pri response_mode=quotes MUSÍ byť claims=[]:
 pri všeobecnej otázke, téme, technológii, zozname ľudí alebo stáži vyber iba
 evidence_quotes: source_id a presný doslovný výňatok, ktorý odpovedá na otázku.
+Pri výbere zachovaj negácie, podmienky a výhrady zdroja. Nevytrhni tvrdenie z vety,
+ktorá ho obmedzuje alebo spochybňuje.
 Nevypĺňaj pri nich názvy prác, roky, tagy ani iné nežiadané metadáta do claims.
 Pri requirements.content=true otázka žiada obsah práce, nie iba jej identifikáciu.
 Vyber opisujúce evidence_quotes z content_passages zodpovedajúcej osoby, typu a roku.
@@ -65,6 +69,11 @@ aplikácia uvedie, čo chýba. no_answer znamená žiadnu odpoveď a musí mať
 Pri zaznamenanom stave uprednostni konkrétnu implementáciu, natrénovanie modelu a vykonané
 vyhodnotenie pred konzultáciami alebo samotným písaním textu. Ani priebežný stav nevydávaj
 za splnenie všetkých cieľov alebo za konečné výsledky práce.
+Ak sú dodané content_candidates, vyber quotes iba z ich textu. Každý kandidát uvádza
+osobu, typ práce, rok, kind (assignment/task/progress/other) a facets, ktoré dokladá.
+Pokry každú dostupnú požadovanú časť requirements.facets za každú osobu. Pre chýbajúcu časť
+nevymýšľaj dôkaz: aplikácia ju výslovne označí. Čiastočná doložená odpoveď má status=ok.
+Pri výsledkoch použi skutočný zaznamenaný stav; úlohy a plánované experimenty nie sú výsledky.
 Každý výňatok musí byť JEDEN súvislý úsek z konkrétneho zdroja vrátane Markdownu.
 Nespájaj oddelené pasáže cez vynechané riadky; použi viac samostatných kratších quotes.
 Vyber iba minimum potrebné na zodpovedanie otázky. Výňatok nesmieš parafrázovať ani meniť čísla.
@@ -127,7 +136,8 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
         issues.append('missing_required_evidence')
     if requirements.get('response_mode') == 'quotes' and claims:
         issues.append('unrequested_structured_claims')
-    if requirements.get('content') and answer['status'] == 'ok' and not quotes:
+    if (requirements.get('content') and answer['status'] == 'ok' and not quotes
+            and (not requirements.get('facets') or context.get('content_candidates'))):
         issues.append('missing_content_evidence')
     requested = requirements['people']
     types = requirements['work_types']
@@ -159,7 +169,12 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
             matching = [claim for claim in claims if claim['person'] == person]
             for key in ['title', 'year', 'start_year']:
                 if requirements[key] and not any(claim.get(key) is not None for claim in matching):
-                    issues.append('missing_requested_' + key)
+                    if key in requirements.get('available_fields',{}).get(person,[key]):
+                        issues.append('missing_requested_' + key)
+        wanted=[key for key in ['title','year','start_year'] if requirements[key]]
+        if (wanted and not quotes and not any(key in fields for fields in requirements.get('available_fields',{}).values()
+                                             for key in wanted)):
+            issues.append('no_requested_fact_supported')
     if answer['status'] == 'no_answer' and (requirements['academic'] or requirements['start_year']) and requested:
         # Reject an unsupported refusal only if every requested academic tuple
         # is explicitly present. Never invent a value to fill an absent field.
@@ -174,6 +189,10 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
                 and (not requirements['start_year'] or start_supported) and not requirements['unsupported_year']
                 and (not requirements.get('content') or context.get('content_passages'))):
             issues.append('rejected_supported_academic_answer')
+        if (not requirements.get('content') and not requirements['unsupported_year'] and any(
+                requirements[key] and key in fields for fields in requirements.get('available_fields',{}).values()
+                for key in ['title','year','start_year'])):
+            issues.append('rejected_supported_partial_fact')
     if answer['status'] == 'no_answer' and requirements.get('content'):
         query = normalized(context['question'])
         overview = re.search(r'\b(?:comu(?:\s+sa)?\s+ven(?:uj|ov)\w*|o\s+com|opis\w*|popis\w*|obsah\w*|zhrn\w*)\b', query)
@@ -193,6 +212,10 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
         elif source.get('source_url') not in cited:
             issues.append('missing_quote_source')
         if source:
+            if context.get('candidate_only') and not any(c['source_id']==quote['source_id'] and
+                    ' '.join(quote['quote'].split()) in ' '.join(c['text'].split())
+                    for c in context.get('content_candidates', [])):
+                issues.append('quote_outside_delivered_candidates')
             if requirements.get('content'):
                 titles = [row.get(key) for row in facts if row.get('source_id') == quote['source_id']
                           for key in ('title', 'person', 'document_author')]
@@ -218,15 +241,16 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
     if requirements.get('content') and answer['status'] == 'ok':
         for person in requested:
             passages = [p for p in context.get('content_passages', []) if p.get('person') == person]
-            if requirements['response_mode'] == 'mixed':
+            if requirements['response_mode'] == 'mixed' and any(claim.get('person')==person for claim in claims):
                 passages = [p for p in passages if any(claim.get('person') == person and
                     all(claim.get(key) is None or claim[key] == p.get(key) for key in ['work_type', 'year'])
                     for claim in claims)]
-            if not any(quote['source_id'] == p['source_id'] and
+            eligible = any(c.get('person')==person for c in context.get('content_candidates', []))
+            if (not requirements.get('facets') or eligible) and not any(quote['source_id'] == p['source_id'] and
                        ' '.join(quote['quote'].split()) in ' '.join(p['text'].split())
                        for p in passages for quote in quotes):
                 issues.append('missing_bound_content_evidence')
-            if requirements.get('distinguish_progress'):
+            if requirements.get('distinguish_progress') and not requirements.get('facets'):
                 bound_quotes = [quote for quote in quotes if any(
                     quote['source_id'] == p['source_id'] and
                     ' '.join(quote['quote'].split()) in ' '.join(p['text'].split()) for p in passages)]
@@ -235,7 +259,15 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
                 for kind in requirements.get('content_roles', {}).get(person, []):
                     if kind not in kinds:
                         issues.append('missing_' + kind + '_evidence')
-    if (requirements.get('distinguish_progress') and answer['status'] == 'no_answer' and requested
+    if requirements.get('facets'):
+        available={(c['person'],facet) for c in context.get('content_candidates',[]) for facet in c['facets']}
+        if answer['status']=='ok':
+            for person,facet in missing_facets(answer,context):
+                if (person,facet) in available:
+                    issues.append('missing_'+('progress' if facet=='results' else facet)+'_evidence')
+        elif answer['status']=='no_answer' and available:
+            issues.append('rejected_supported_content_comparison')
+    if (requirements.get('distinguish_progress') and not requirements.get('facets') and answer['status'] == 'no_answer' and requested
             and all(any(p.get('person') == person for p in context.get('content_passages', [])) for person in requested)):
         issues.append('rejected_supported_content_comparison')
     if not citation_check(answer, allowed_urls(context))['citations_ok']:
@@ -333,11 +365,29 @@ def comparison_lines(answer: dict, context: dict) -> list[str]:
 def render_rag_answer(answer: dict, context: dict) -> str:
     lines = []
     if answer['status'] == 'ok':
+        req=context.get('requirements',{})
         for claim in answer.get('claims', []):
-            lines.append(fact_sentences(claim, context.get('question', '')))
+            if req.get('facets'):
+                # Facet headings already identify the scoped work/year. Extra
+                # valid metadata selected by the model is not part of the answer.
+                fields=[key for key in ['title','year','start_year'] if req.get(key)]
+                if not req.get('title') and not req.get('start_year') and req.get('years'):
+                    continue
+                displayed={key:value for key,value in claim.items() if key in ['person','work_type',*fields]}
+                if not any(displayed.get(key) is not None for key in fields):
+                    continue
+                lines.append(fact_sentences(displayed,context.get('question','')))
+            else:
+                lines.append(fact_sentences(claim, context.get('question', '')))
+        for person in req.get('people',[]):
+            for key,label in [('title','Názov práce'),('year','Rok práce'),('start_year','Začiatok štúdia')]:
+                if req.get(key) and key not in req.get('available_fields',{}).get(person,[key]):
+                    lines.append(f'{person} — {label.lower()} nie je v poskytnutých podkladoch doložený.')
         quotes = answer.get('evidence_quotes', [])
         style = context.get('requirements', {}).get('content_style')
-        if quotes and context.get('requirements', {}).get('distinguish_progress'):
+        if context.get('requirements', {}).get('facets'):
+            lines.extend(facet_lines(answer,context))
+        elif quotes and context.get('requirements', {}).get('distinguish_progress'):
             lines.extend(comparison_lines(answer, context))
         elif quotes and style in {'overview', 'assignment'}:
             # Bind each paragraph to the same source/person/work as the checked
@@ -371,7 +421,10 @@ def render_rag_answer(answer: dict, context: dict) -> str:
         if answer.get('count') is not None:
             lines.append(f'Počet: {answer["count"]}.')
     elif answer['status'] == 'no_answer':
-        lines.append(no_answer_text(context))
+        if context.get('requirements', {}).get('facets') and context['requirements']['people']:
+            lines.extend(facet_lines(answer,context))
+        else:
+            lines.append(no_answer_text(context))
     elif answer['status'] == 'ambiguous':
         lines.append('Údaje sú nejednoznačné; požiadavku treba spresniť.')
     else:
@@ -384,6 +437,18 @@ def render_rag_answer(answer: dict, context: dict) -> str:
 
 def generate_verified_rag(question: str, payload: dict, complete) -> dict:
     context = rag_evidence_context(question, payload)
+    if payload.get('status')=='ambiguous' or context.get('clarification'):
+        return {'status':'ambiguous','reason':payload.get('reason') or 'ambiguous_request',
+                'accepted_for_delivery':True,'answer':context.get('clarification') or 'Spresni osobu, typ práce alebo rok.',
+                'claims':[],'source_urls':[],'attempts':[],'complete':False,'truncated':False,
+                'truncation_reasons':[],'index_version':context.get('index_version')}
+    if (context['requirements'].get('facets') and context['requirements']['response_mode']=='quotes'
+            and not context.get('content_candidates') and context['requirements']['people']):
+        empty={'status':'no_answer','claims':[],'evidence_quotes':[],
+               'source_urls':[s['source_url'] for s in context['sources'] if s.get('source_url')]}
+        return {**empty,'reason':'requested_aspects_not_found','answer':render_rag_answer(empty,context),
+                'accepted_for_delivery':True,'attempts':[],'answer_coverage':answer_coverage(empty,context),
+                **{k:context.get(k) for k in ['index_version','complete','truncated','truncation_reasons']}}
     if not payload.get('sources'):
         status = 'error' if payload.get('status') == 'error' else 'no_answer'
         empty = {'status': status, 'claims': [], 'evidence_quotes': [], 'source_urls': []}
@@ -395,11 +460,18 @@ def generate_verified_rag(question: str, payload: dict, complete) -> dict:
     sent = {key: context[key] for key in ['status', 'reason', 'complete', 'truncated',
         'truncation_reasons', 'index_version', 'sources', 'facts', 'requirements']}
     if context['requirements'].get('content'):
-        sent['content_passages'] = context['content_passages']
+        if context['requirements'].get('facets'):
+            sent['content_candidates']=context['content_candidates']
+            context['candidate_only']=True
+        else:
+            sent['content_passages'] = context['content_passages']
     elif not context['requirements']['academic']:
         sent['context'] = payload.get('context', '')
     if context['requirements']['response_mode'] == 'quotes':
         sent.pop('facts')
+    if context['requirements'].get('facets'):
+        fit_candidates(sent,context,payload.get('context_budget',{}).get('max_tokens',12000),
+                       os.getenv('RAG_LLM_TOKEN_ENCODING','cl100k_base'))
     messages = [{'role': 'system', 'content': PROMPT + '\nSchéma: ' + json.dumps(RagAnswer.model_json_schema(), ensure_ascii=False)},
                 {'role': 'user', 'content': json.dumps({'question': question, 'context': sent}, ensure_ascii=False)}]
     attempts = []
@@ -420,6 +492,12 @@ def generate_verified_rag(question: str, payload: dict, complete) -> dict:
                     'answer': 'Jazykový model nie je dostupný; overenú odpoveď sa nepodarilo pripraviť.', 'attempts': attempts}
         record['issues'] = issues
         if not issues:
+            if context['requirements'].get('facets') and missing_facets(answer,context):
+                # Sources inspected for a missing requested aspect remain
+                # traceable, even when only the other person's quotes exist.
+                extra=[s['source_url'] for s in context['sources'] if s.get('title') in
+                       {person for person,_ in missing_facets(answer,context)} and s.get('source_url')]
+                answer={**answer,'source_urls':list(dict.fromkeys([*answer['source_urls'],*extra]))}
             text = render_rag_answer(answer, context)
             return {'status': answer['status'], 'reason': None, 'accepted_for_delivery': True,
                     'answer': text, 'source_urls': answer['source_urls'], 'claims': answer['claims'],

tools/answers/answer_delivery.py

@@ -12,12 +12,20 @@ from tools.answer_evaluation.graph_context import graph_context
 from tools.answer_evaluation.checked_generation import generate_checked_answer
 from tools.answer_evaluation.verified_rag import generate_verified_rag
 from tools.answer_evaluation.rag_evidence import expand_comparison_sources
+from tools.answer_evaluation.rag_preparation import prepare_named_content
+from tools.retrieval.question_constraints import unsupported_constraint, constraint_message
 from tools.sync.index_state import index_snapshot
 from tools.answers.model_client import complete_with_configured_model
+from tools.answers.answer_errors import public_failure
 
 _capacity = threading.BoundedSemaphore(2)
 
 
+def retrieval_failure(engine: str) -> dict:
+    return public_failure({'reason':'retrieval_unavailable','engine':engine,'index_version':None,
+                           'complete':False,'truncated':False,'truncation_reasons':[]})
+
+
 def deliver_answer(question: str, *, engine: str = 'rag', limit: int = 5,
                    complete=None, db_file: Path = DB_FILE, include_diagnostics: bool = False) -> dict:
     if not _capacity.acquire(blocking=False):
@@ -25,15 +33,32 @@ def deliver_answer(question: str, *, engine: str = 'rag', limit: int = 5,
                 'accepted_for_delivery': False, 'answer': 'Agent práve spracúva iné odpovede. Skúste to o chvíľu.'}
     try:
         model = complete or complete_with_configured_model
+        unsupported=unsupported_constraint(question)
+        if unsupported:
+            return {'status':'ambiguous','reason':'unsupported_constraint','accepted_for_delivery':True,
+                    'answer':constraint_message(unsupported),'claims':[],'source_urls':[],
+                    'index_version':None,'complete':False,'truncated':False,'truncation_reasons':[], 'engine':engine}
         if engine == 'rag':
-            with index_snapshot(db_file.parent, db_file=db_file):
-                retrieved = build_rag_context(db_file, question, limit=limit)
-                retrieved = expand_comparison_sources(db_file, question, retrieved)
-                retrieved = compact_rag_payload(retrieved)
+            try:
+                with index_snapshot(db_file.parent, db_file=db_file):
+                    retrieved = build_rag_context(db_file, question, limit=limit)
+                    retrieved = expand_comparison_sources(db_file, question, retrieved)
+                    retrieved = prepare_named_content(db_file, question, retrieved)
+                    retrieved = compact_rag_payload(retrieved)
+            except Exception:
+                return retrieval_failure(engine)
             result = generate_verified_rag(question, retrieved, model)
             context = retrieved
         elif engine == 'graphrag':
-            context = graph_context(search_graph(question))
+            try:
+                context = graph_context(search_graph(question))
+            except Exception:
+                return retrieval_failure(engine)
+            if context.get('clarification'):
+                return {'status':'ambiguous','reason':context.get('reason'),'accepted_for_delivery':True,
+                        'answer':context['clarification'],'claims':[],'source_urls':[], 'engine':engine,
+                        'complete':False,'truncated':False,'truncation_reasons':[],
+                        'index_version':context.get('index_version')}
             prompt = (Path(__file__).resolve().parents[1] / 'answer_evaluation/graph_answer_prompt.txt').read_text()
             generated = generate_checked_answer(question, context, prompt, model)
             if generated.get('accepted_for_delivery'):
@@ -42,8 +67,10 @@ def deliver_answer(question: str, *, engine: str = 'rag', limit: int = 5,
                     'truncated': context.get('truncated'), 'truncation_reasons': context.get('truncation_reasons'),
                     'attempts': generated['attempts']}
             else:
-                result = {'status': 'error', 'reason': 'answer_rejected', 'accepted_for_delivery': False,
-                    'answer': 'Odpoveď neprešla kontrolou podkladov; neoverené údaje sa nezobrazia.',
+                reason=generated.get('reason','answer_rejected')
+                result = {'status': 'error', 'reason': reason, 'accepted_for_delivery': False,
+                    'answer': ('Jazykový model nie je dostupný. Skús otázku neskôr.' if reason=='model_unavailable' else
+                               'Odpoveď neprešla kontrolou podkladov; neoverené údaje sa nezobrazia.'),
                     'index_version': context.get('index_version'), 'attempts': generated['attempts']}
         else:
             raise ValueError('Unsupported answer engine')

tools/answers/answer_errors.py

@@ -0,0 +1,18 @@
+"""Public failure wording; transport/model exception text is never a UI message."""
+MESSAGES={
+    'model_unavailable':'Jazykový model momentálne nie je dostupný. Skús otázku neskôr.',
+    'retrieval_unavailable':'Vyhľadávací index momentálne nie je dostupný. Skús otázku neskôr.',
+    'answer_capacity_exceeded':'Agent práve spracúva iné otázky. Skús to o chvíľu.',
+    'answer_rejected':'Odpoveď sa nepodarilo spoľahlivo overiť podľa podkladov. '
+                      'Skús spresniť osobu, typ práce alebo rok; neoverené údaje sa nezobrazia.',
+    'answer_unavailable':'Overenú odpoveď sa momentálne nepodarilo pripraviť. Skús to neskôr.',
+}
+
+
+def public_failure(result: dict) -> dict:
+    reason=result.get('reason')
+    if not isinstance(reason,str) or reason not in MESSAGES:
+        reason='answer_unavailable'
+    return {'index_version':None,'complete':False,'truncated':False,'truncation_reasons':[],
+            **result,'status':'error','reason':reason,'accepted_for_delivery':False,
+            'answer':MESSAGES[reason],'claims':[],'source_urls':[]}

tools/graphrag/evaluation/reproducibility.py

@@ -46,6 +46,8 @@ def content_fingerprints(questions: Path, root: Path = ROOT) -> dict[str, Any]:
         'tools/answer_evaluation/render_answer.py', 'tools/answer_evaluation/graph_context.py',
         'tools/answer_evaluation/presentation.py',
         'tools/answer_evaluation/rag_evidence.py',
+        'tools/answer_evaluation/question_intent.py', 'tools/answer_evaluation/content_selection.py',
+        'tools/answer_evaluation/rag_preparation.py', 'tools/answer_evaluation/answer_coverage.py',
         'tools/answer_evaluation/graph_answer_prompt.txt',
         'tools/rag/evaluation/rag_metrics.py', 'tools/rag/evaluation/rag_runner.py'])
     files = sorted(set(files))

tools/graphrag/graph_search.py

@@ -8,6 +8,7 @@ from typing import Any, Iterable
 from tools.retrieval.publication import DEFAULT_PUBLISHED_ONLY, PUBLICATION_POLICY_VERSION
 from .publication import public_query
 from tools.retrieval.retrieval_contract import result_fields, source_references, error_result
+from tools.retrieval.question_constraints import unsupported_constraint, constraint_message
 
 
 GRAPH_SCOPE = "zpwiki"
@@ -2334,6 +2335,12 @@ def search_graph_with_session(
         raise ValueError("max_entities musí byť aspoň 1")
 
     constraints = extract_query_constraints(clean_query)
+    unsupported = unsupported_constraint(clean_query)
+    if unsupported:
+        return {**result_fields(status='ambiguous',sources=[],reason='unsupported_constraint',complete=False),
+                'query':clean_query,'constraints':constraints,'clarification':constraint_message(unsupported),
+                'facts':[],'paths':[],'documents':[],'aggregation':None,'resolved_entities':[],
+                'checked_entities':[],'rejection_evidence':[],'strategies':[],'reasons':['unsupported_constraint']}
     catalog = run_rows(session, CATALOG_QUERY, scope=GRAPH_SCOPE, published_only=published_only)
     constraints['published_only'] = published_only
     # Explicit tag filters are resolved separately. Incidental names such as the

tools/rag/llm_context.py

@@ -11,7 +11,7 @@ from tools.retrieval.text_budget import split_to_fit
 SOURCE_FIELDS = ('source_id', 'title', 'author', 'document_path', 'source_url', 'published',
                  'categories', 'tags', 'section')
 FIELDS = ('status', 'reason', 'query', 'engine', 'strategies', 'instructions', 'answer_format',
-          'complete', 'truncated', 'truncation_reasons', 'index_version', 'source_count')
+          'complete', 'truncated', 'truncation_reasons', 'index_version', 'source_count', 'clarification', 'named_subjects')
 
 
 def compact_rag_payload(data: dict, *, max_tokens: int | None = None,

tools/rag/rag_utils.py

@@ -709,6 +709,13 @@ def build_rag_context(
     published_only: bool = DEFAULT_PUBLISHED_ONLY,
     max_per_document: int = 1,
 ) -> dict[str, Any]:
+    from tools.retrieval.question_constraints import unsupported_constraint, constraint_message
+    unsupported=unsupported_constraint(query)
+    if unsupported:
+        return {**result_fields(status='ambiguous',sources=[],reason='unsupported_constraint',complete=False),
+                'query':query,'engine':'hybrid','strategies':[],'source_count':0,'sources':[],
+                'context':'','clarification':constraint_message(unsupported),
+                'instructions':RAG_INSTRUCTIONS,'answer_format':ANSWER_FORMAT}
     response = search_database(
         db_path,
         query,

tools/retrieval/question_constraints.py

@@ -0,0 +1,42 @@
+"""Fail closed for recognized query operators without an implemented meaning."""
+import re
+import unicodedata
+
+
+def requested_years(question: str) -> list[int]:
+    """An unquoted four-digit year remains a condition without 'v roku'."""
+    text=re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', question)
+    return list(dict.fromkeys(int(value) for value in re.findall(r'\b\d{4}\b',text)))
+
+
+def unsupported_constraint(question: str) -> str | None:
+    text = ''.join(c for c in unicodedata.normalize('NFKD', question).casefold()
+                   if not unicodedata.combining(c))
+    # Do not treat words inside quoted academic titles as query operators.
+    text = re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', text)
+    if (re.search(r'\b(?:rok\w*|datum\w*|kedy)\b',text) and
+        re.search(r'\b(?:obhajob\w*|obhajil\w*|dokonc\w*|ukoncil\w*|zacal\w* pracovat)\b',text) and
+        re.search(r'\b(?:prac\w*|projekt\w*)\b',text)):
+        return 'work_event_date'
+    if re.search(r'\b(?:od|do|pred|po|medzi)\s+(?:rok(?:om|mi|u)?\s+)?\d{4}\b|\b\d{4}\s*[-–]\s*\d{4}\b', text):
+        return 'year_range'
+    if re.search(r'\b(?:aspon|najviac|viac ako|menej ako|minimalne|maximalne)\s+\d+', text):
+        return 'quantity_condition'
+    if re.search(r'\b(?:okrem|s vynimkou)\b|\b(?:nema|nemaju)\s+(?:tag|klucov|kategori|prac|projekt)|'
+                 r'\bbez\s+(?:tagu|klucoveho slova|kategorie)\b', text):
+        return 'exclusion'
+    if re.search(r'\b(?:tag\w*|klucov\w*\s+slov\w*)\s+\w+\s+alebo\s+\w+', text):
+        return 'alternative_tags'
+    return None
+
+
+def constraint_message(code: str) -> str:
+    return {
+        'work_event_date':'Rok uvedený pri práci nepreukazuje dátum jej začatia, dokončenia alebo obhajoby. '
+                         'Chceš zistiť rok evidovaný pri práci?',
+        'year_range': 'Časovú podmienku neviem spoľahlivo vyhodnotiť ako interval. Uveď konkrétny rok práce.',
+        'quantity_condition': 'Podmienku na minimálny alebo maximálny počet neviem spoľahlivo vyhodnotiť. '
+                              'Požiadaj najprv o konkrétny zoznam alebo počet.',
+        'exclusion': 'Vylučovaciu podmienku neviem spoľahlivo vyhodnotiť. Uveď, ktoré vlastnosti majú výsledky spĺňať.',
+        'alternative_tags': 'Uveď, či požaduješ oba tagy zároveň, alebo vyhľadávanie pre každý tag samostatne.',
+    }[code]