Presnejšie RAG 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 1.1.0) a over správanie v novom chate.
+novými ZIPmi z `packages/claude/dist/` (verzia 1.2.0) a over správanie v novom chate.
 
 ### Nahratie Skills do webového Claude
 
@@ -1475,6 +1475,23 @@ staré ploché cesty `data/nazov.json` sa presmerujú do príslušnej oblasti.
 
 ### Širšie overenie kontrolovaného RAG
 
+`complete`, `truncated` a `truncation_reasons` zostávajú technickým opisom
+vyhľadávania. Overená odpoveď navyše obsahuje `answer_coverage` so scope, status,
+warnings a limitations. Status `supported` znamená doložené tvrdenia/výňatky,
+nie úplné prehľadanie korpusu. Bežný výber kandidátov a zdrojov preto nepridáva
+automatické upozornenie k názvu, roku alebo krátkemu opisu. Obmedzený zoznam
+záznamov a orezané citované zadanie dostanú konkrétne upozornenie. Pri nenájdenej
+informácii sa zachová opatrné tvrdenie o dostupných podkladoch. Žiadny limit sa
+kvôli tomu nezvyšuje ani sa neprepisuje na `complete=true`.
+
+Otázka na stručný obsah pri zobrazení vyberá najviac dva krátke overené výňatky
+na osobu (do 600 znakov a dvoch bodov na osobu), bez voľnej modelovej parafrázy.
+Skrátenie robí renderer z už overených citátov; dlhší platný citát nie je dôvod
+na odmietnutie faktov ani ďalšie modelové volanie. Pri neoddeliteľnom technickom
+reťazci má zachovanie dôkazu prednosť pred limitom stručnosti. Otázka na zadanie
+môže zobraziť všetky doložené body zadania. Citácie a kontroly ich doslovnosti
+sa uplatňujú v oboch režimoch.
+
 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

packages/claude/client.py

@@ -163,7 +163,14 @@ def query_backend(query: str, engine: str, *, mode='answer', limit=5, config=Non
                     or not all(isinstance(x, str) for x in value['source_urls'])
                     or not isinstance(value.get('claims'), list)):
                 raise ClientError('invalid_backend_response')
-            fields |= {'answer', 'source_urls', 'claims', 'count', 'accepted_for_delivery'}
+            if 'answer_coverage' in value:
+                coverage = value['answer_coverage']
+                if (not isinstance(coverage, dict) or coverage.get('scope') not in ('fact','overview','assignment','list','count')
+                        or coverage.get('status') not in ('supported','partial','unknown','not_found','unavailable')
+                        or any(not isinstance(coverage.get(key), list) or
+                               any(not isinstance(item, str) for item in coverage[key]) for key in ['warnings','limitations'])):
+                    raise ClientError('invalid_backend_response')
+            fields |= {'answer', 'source_urls', 'claims', 'count', 'accepted_for_delivery', 'answer_coverage'}
         else:
             if not isinstance(value.get('sources'), list):
                 raise ClientError('invalid_backend_response')

packages/claude/dist/packages.json

@@ -3,12 +3,12 @@
     {
       "archive": "zpwiki-rag.zip",
       "name": "rag",
-      "sha256": "3382847549d88c6b3c25dc29cc95c9d8ceb8df1b3558680e772374a06b1a04c8"
+      "sha256": "8f0c51e27dc68a1e3ee845f2cfe5987e51714d4408f98c7d275861dc1e22953f"
     },
     {
       "archive": "zpwiki-graphrag.zip",
       "name": "graphrag",
-      "sha256": "f1090fa4e22fb1cd4ebe3cc1a5c9b0c7d8fb7a57a8057713223007060df7f59a"
+      "sha256": "aa65a97211f48fa0352a98e50c6ca6a243e1d23105c0ff144c7f360ad192c376"
     }
   ],
   "schema_version": 1

packages/claude/mcp_server.py

@@ -25,6 +25,10 @@ ANSWER_PRESENTATION_POLICY = (
     'document_author means author of the source document; it does not establish thesis authorship, '
     'supervision or a maintainer role. Do not guess those roles. '
     'complete and truncated describe retrieval coverage, not the truth of each returned fact. '
+    'For RAG, answer_coverage separately describes the delivered answer scope. supported means '
+    'evidence-backed, not exhaustive. Do not append a generic incompleteness warning merely '
+    'because retrieval selected top candidates/results. Preserve the specific notices already '
+    'included in answer for limited lists or clipped source text. '
     'An incomplete listing can contain supported facts, but cannot establish uniqueness or absence '
     'of other records. Retain its caveat; do not dismiss it using earlier answers. '
     'For no_answer, return only the supplied absence-of-evidence answer and any supplied citations. '

packages/claude/rag/SKILL.md

@@ -22,13 +22,17 @@ Use this text-search tool rather than GraphRAG when the user explicitly selects
    never reproduce rejected text or invent a replacement. For `ambiguous`, ask the user
    to clarify. For `no_answer`, report the absence of evidence without claiming the fact is false.
 5. Respect `complete`, `truncated`, `truncation_reasons` and `index_version`.
+   These describe retrieval, not whether the selected answer is supported. `answer_coverage`
+   separately describes the answer scope: `supported` means evidence-backed, never exhaustive.
+   Present the backend's specific notices unchanged. Do not append a generic incompleteness
+   warning to a supported fact or short description just because retrieval selected top results.
    Never describe an incomplete or unknown-completeness result as an exhaustive list.
    Do not combine different index versions as if they represented one snapshot.
 
 Presentation boundaries:
 - Use the current tool response for the current question. Do not append factual material
   or reassurances from previous chat answers.
-- An incomplete listing can contain supported facts. Preserve its caveat without implying
+- An incomplete listing can contain supported facts. Preserve any supplied caveat without implying
   the returned title/year is wrong or that it is the only matching record. Earlier answers
   cannot waive the current response's coverage limits.
 - `document_author` identifies the source document's author. Do not infer thesis authorship,

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 = '1.1.0'
+PACKAGE_VERSION = '1.2.0'
 
 
 def json_bytes(value):

test/test_claude_packages.py

@@ -132,6 +132,25 @@ def test_presentation_instructions_never_escape_and_evidence_is_preserved(backen
     assert backend['checked'] == original
 
 
+def test_answer_coverage_is_forwarded_without_overwriting_retrieval_limits(backend):
+    coverage = {'scope': 'fact', 'status': 'supported', 'warnings': [], 'limitations': ['result_limit']}
+    backend['checked']['answer_coverage'] = coverage
+    result = client.query_backend('Názov práce?', 'rag')
+    assert result['answer_coverage'] == coverage
+    assert result['complete'] is False and result['truncated'] is True
+    assert 'presentation_instructions' not in result
+
+
+@pytest.mark.parametrize('coverage', [None, [], {'status': 'complete'},
+    {'scope': 'fact', 'status': 'supported', 'warnings': 'none', 'limitations': []},
+    {'scope': [], 'status': 'supported', 'warnings': [], 'limitations': []},
+    {'scope': 'fact', 'status': {}, 'warnings': [], 'limitations': []}])
+def test_invalid_answer_coverage_is_not_delivered(backend, coverage):
+    backend['checked']['answer_coverage'] = coverage
+    result = client.query_backend('Názov práce?', 'rag')
+    assert result['reason'] == 'invalid_backend_response' and not result['accepted_for_delivery']
+
+
 @pytest.mark.parametrize('code,reason', [(401, 'unauthorized'), (403, 'forbidden'),
     (503, 'backend_unavailable'), (302, 'redirect_not_allowed')])
 def test_transport_errors_and_redirects_do_not_disclose_payload(backend, code, reason):

test/test_rag_answer_coverage.py

@@ -0,0 +1,198 @@
+"""Technical retrieval limits must not become false claims about answer quality."""
+import copy
+import json
+
+import pytest
+
+from tools.answer_evaluation.answer_coverage import answer_coverage
+from tools.answer_evaluation.rag_evidence import rag_evidence_context
+from tools.answer_evaluation.verified_rag import generate_verified_rag, rag_answer_issues, render_rag_answer
+from tools.rag.rag_utils import build_context_text
+
+
+URL = 'https://example.test/eva'
+GOALS = '1. Preskúmajte metódy vyhľadávania.\n2. Implementujte vyhľadávač.\n3. Vyhodnoťte presnosť.\n4. Navrhnite zlepšenia.'
+
+
+@pytest.fixture
+def payload():
+    sources = [{'source_id': 'S1', 'title': 'Eva Nová', 'document_path': 'pages/students/2020/eva_nova/README.md',
+        'source_url': URL, 'text': 'Sekcia: Diplomová práca 2026\nNázov: Vyhľadávanie informácií.\nZadanie:\n' + GOALS}]
+    return {'status': 'ok', 'reason': None, 'complete': False, 'truncated': True,
+            'truncation_reasons': ['candidate_limit', 'per_document_limit', 'result_limit'],
+            'index_version': {'version': 'fixture'}, 'sources': sources, 'context': build_context_text(sources)}
+
+
+def answer():
+    return {'status': 'ok', 'answer': '', 'claims': [{'person': 'Eva Nová', 'work_type': 'diploma_thesis',
+        'title': 'Vyhľadávanie informácií', 'year': 2026}], 'evidence_quotes': [], 'source_urls': [URL]}
+
+
+def quote(text):
+    return {'status': 'ok', 'answer': '', 'claims': [], 'source_urls': [URL],
+            'evidence_quotes': [{'source_id': 'S1', 'quote': text}]}
+
+
+@pytest.mark.parametrize('complete,truncated,reasons', [
+    (False, True, ['candidate_limit', 'per_document_limit', 'result_limit']),
+    (None, False, []), (False, True, ['llm_context_budget'])])
+def test_supported_fact_is_not_an_incomplete_answer(payload, complete, truncated, reasons):
+    payload.update(complete=complete, truncated=truncated, truncation_reasons=reasons)
+    c = rag_evidence_context('Aký je názov a rok diplomovej práce Evy Novej?', payload)
+    before = copy.deepcopy(c)
+    coverage = answer_coverage(answer(), c)
+    assert coverage['status'] == 'supported' and coverage['warnings'] == []
+    text = render_rag_answer(answer(), c)
+    assert '2026' in text and URL in text and 'neúpln' not in text
+    assert c == before  # Never fake complete=true or remove diagnostic limits.
+
+
+@pytest.mark.parametrize('question', ['Vypíš všetky práce Evy Novej.', 'Aké práce má Eva Nová?',
+                                     'Uveď kompletný zoznam prác Evy Novej.',
+                                     'Vypíš všetky bakalárske a diplomové práce Evy Novej.'])
+def test_limited_record_list_still_warns(payload, question):
+    c = rag_evidence_context(question, payload)
+    coverage = answer_coverage(answer(), c)
+    assert coverage['scope'] == 'list' and coverage['status'] == 'partial'
+    assert 'Uvedený zoznam nemusí obsahovať všetky záznamy.' in render_rag_answer(answer(), c)
+
+
+def test_explicitly_complete_retrieval_does_not_add_a_list_warning(payload):
+    payload.update(complete=True, truncated=False, truncation_reasons=[])
+    c = rag_evidence_context('Vypíš všetky práce Evy Novej.', payload)
+    assert answer_coverage(answer(), c)['warnings'] == []
+
+
+def test_quoted_title_is_not_an_exhaustive_list_request(payload):
+    c = rag_evidence_context('Aký je rok diplomovej práce „Všetky práce v katalógu“?', payload)
+    assert answer_coverage(answer(), c)['scope'] == 'fact'
+
+
+def test_unknown_list_coverage_is_not_silently_exhaustive(payload):
+    payload.update(complete=None, truncated=False, truncation_reasons=[])
+    c = rag_evidence_context('Aké práce má Eva Nová?', payload)
+    assert answer_coverage(answer(), c)['warnings'] == ['list_may_be_incomplete']
+
+
+def test_legacy_context_without_limit_metadata_remains_usable(payload):
+    payload.pop('truncation_reasons'); payload.pop('complete'); payload.pop('truncated')
+    c = rag_evidence_context('Aký je názov diplomovej práce Evy Novej?', payload)
+    assert answer_coverage(answer(), c)['warnings'] == []
+    assert 'Vyhľadávanie informácií' in render_rag_answer(answer(), c)
+
+
+def test_regular_assignment_can_show_all_supplied_goals_without_global_warning(payload):
+    c = rag_evidence_context('Aké je zadanie diplomovej práce Evy Novej?', payload)
+    a = quote(GOALS)
+    assert rag_answer_issues(a, c) == []
+    text = render_rag_answer(a, c)
+    assert GOALS in text and 'neúpln' not in text and 'Zo zdroja' not in text
+    assert answer_coverage(a, c)['scope'] == 'assignment'
+
+
+def test_clipped_assignment_gets_specific_warning(payload):
+    payload['sources'][0]['text'] += '\n[OBSAH ZDROJA OREZANÝ]'
+    payload['context'] = build_context_text(payload['sources'])
+    c = rag_evidence_context('Aké je zadanie diplomovej práce Evy Novej?', payload)
+    a = quote('1. Preskúmajte metódy vyhľadávania.')
+    coverage = answer_coverage(a, c)
+    assert coverage['status'] == 'partial' and coverage['warnings'] == ['assignment_text_clipped']
+    assert 'uvedené body nemusia tvoriť celé zadanie' in render_rag_answer(a, c)
+
+
+def test_request_for_all_assignment_items_preserves_uncertainty(payload):
+    c = rag_evidence_context('Uveď všetky body zadania diplomovej práce Evy Novej.', payload)
+    assert answer_coverage(quote(GOALS), c)['warnings'] == ['assignment_completeness_unknown']
+
+
+@pytest.mark.parametrize('budget', [False, True])
+def test_no_answer_is_bounded_and_only_reports_actual_context_cut(payload, budget):
+    payload['truncation_reasons'] = ['llm_context_budget'] if budget else ['result_limit']
+    c = rag_evidence_context('Aké je zadanie diplomovej práce Evy Novej v roku 1999?', payload)
+    a = {'status': 'no_answer', 'claims': [], 'evidence_quotes': [], 'source_urls': []}
+    text = render_rag_answer(a, c)
+    assert 'nepodarilo spoľahlivo nájsť' in text and 'Výpis je neúplný' not in text
+    assert ('Časť podkladov bola pri príprave kontextu skrátená.' in text) == budget
+    assert answer_coverage(a, c)['status'] == 'not_found'
+
+
+def test_irrelevant_dropped_source_does_not_make_cited_assignment_partial(payload):
+    payload['sources'].append({'source_id': 'S2', 'title': 'Iný dokument', 'source_url': 'https://example.test/other',
+        'document_path': 'pages/topics/other/README.md', 'text': 'Iný obsah\n[OBSAH ZDROJA OREZANÝ]'})
+    payload['context'] = build_context_text(payload['sources'])
+    payload['truncation_reasons'].append('llm_context_budget')
+    c = rag_evidence_context('Aké je zadanie diplomovej práce Evy Novej?', payload)
+    assert answer_coverage(quote(GOALS), c)['warnings'] == []
+
+
+def test_exact_aggregate_can_be_supported_despite_limited_evidence(payload):
+    c = rag_evidence_context('Koľko prác je evidovaných?', payload)
+    c['aggregation'] = {'count': 2, 'complete': True}
+    a = {'status': 'ok', 'claims': [], 'evidence_quotes': [], 'source_urls': [], 'count': 2}
+    assert answer_coverage(a, c)['status'] == 'supported'
+    c['aggregation']['complete'] = False
+    assert 'count_not_established' in answer_coverage(a, c)['warnings']
+
+
+def test_long_valid_overview_is_shortened_without_another_model_request(payload):
+    question = 'Čomu sa venuje diplomová práca Evy Novej?'
+    short = '1. Preskúmajte metódy vyhľadávania.\n2. Implementujte vyhľadávač.'
+    c = rag_evidence_context(question, payload)
+    assert rag_answer_issues(quote(GOALS), c) == []
+    calls = []
+    def complete(messages):
+        calls.append(messages)
+        return {'choices': [{'message': {'content': json.dumps(quote(GOALS))}}]}
+    result = generate_verified_rag(question, payload, complete)
+    assert result['accepted_for_delivery'] and len(result['attempts']) == len(calls) == 1
+    assert short in result['answer'] and '3.' not in result['answer']
+    assert result['attempts'][0]['parsed_answer']['evidence_quotes'][0]['quote'] == GOALS
+    assert result['complete'] is False and result['truncation_reasons'] == payload['truncation_reasons']
+    assert result['answer_coverage']['status'] == 'supported' and result['answer_coverage']['warnings'] == []
+
+
+def test_short_comparison_can_cover_three_people_instead_of_dropping_one(payload):
+    sources = []
+    quotes = []
+    for index, (name, slug) in enumerate([('Eva Nová', 'eva_nova'), ('Boris Nový', 'boris_novy'),
+                                        ('Cyril Malý', 'cyril_maly')], 1):
+        source = {**payload['sources'][0], 'source_id': f'S{index}', 'title': name,
+                  'source_url': f'https://example.test/{slug}',
+                  'document_path': f'pages/students/2020/{slug}/README.md'}
+        sources.append(source)
+        quotes.append({'source_id': source['source_id'], 'quote': '2. Implementujte vyhľadávač.'})
+    payload['sources'] = sources; payload['context'] = build_context_text(sources)
+    c = rag_evidence_context('Porovnaj obsah diplomových prác osôb Eva Nová, Boris Nový a Cyril Malý.', payload)
+    assert len(c['requirements']['people']) == 3
+    a = {'status': 'ok', 'claims': [], 'evidence_quotes': quotes, 'source_urls': [s['source_url'] for s in sources]}
+    assert rag_answer_issues(a, c) == []
+    a['evidence_quotes'].pop()
+    assert 'missing_bound_content_evidence' in rag_answer_issues(a, c)
+
+
+def test_overview_renderer_only_selects_literal_units_and_preserves_full_evidence(payload):
+    from tools.answer_evaluation.verified_rag import overview_excerpts
+    extra = 'Stav:\n\n- Implementované indexovanie dokumentov.'
+    payload['sources'][0]['text'] += '\n' + extra
+    payload['context'] = build_context_text(payload['sources'])
+    c = rag_evidence_context('Čomu sa venuje diplomová práca Evy Novej?', payload)
+    a = quote(GOALS)
+    a['evidence_quotes'].append({'source_id': 'S1', 'quote': extra})
+    before = copy.deepcopy(a)
+    assert rag_answer_issues(a, c) == []
+    units = overview_excerpts(a, c)
+    assert len(units) == 2 and sum(map(len, units)) <= 600
+    assert all(any(unit in evidence['quote'] for evidence in a['evidence_quotes']) for unit in units)
+    assert a == before
+
+
+def test_long_prose_excerpt_is_selected_without_inventing_a_summary(payload):
+    from tools.answer_evaluation.verified_rag import overview_excerpts
+    prose = 'Práca skúma vyhľadávanie v slovenských dokumentoch. ' * 18
+    payload['sources'][0]['text'] += '\n' + prose
+    payload['context'] = build_context_text(payload['sources'])
+    c = rag_evidence_context('Čomu sa venuje diplomová práca Evy Novej?', payload)
+    a = quote(prose)
+    assert rag_answer_issues(a, c) == []
+    units = overview_excerpts(a, c)
+    assert units and sum(map(len, units)) <= 600 and all(unit in prose for unit in units)

test/test_rag_work_content.py

@@ -153,7 +153,8 @@ def test_correction_sends_real_description_but_no_unrelated_sections(payload):
     text = json.dumps(sent['content_passages'], ensure_ascii=False)
     assert 'ohodnotenia dokumentov' in text and 'Cassandra' not in text and 'medicínskych' not in text
     assert 'missing_content_evidence' in result['attempts'][0]['issues']
-    assert ASSIGNMENT in result['answer'] and 'Výpis je neúplný.' in result['answer']
+    assert ASSIGNMENT in result['answer'] and 'Výpis je neúplný.' not in result['answer']
+    assert result['complete'] is False and result['answer_coverage']['status'] == 'supported'
     assert 'expected_bindings' not in json.dumps(calls)
 
 

test/test_verified_rag.py

@@ -104,7 +104,8 @@ def test_bad_answer_is_repaired_once_and_only_verified_fields_are_rendered(paylo
     assert result['accepted_for_delivery'] and len(calls) == 2
     assert 'Ondrej Megela — bakalárska práca' in result['answer']
     assert 'rok práce: 2020' in result['answer'] and '2099' not in result['answer']
-    assert 'Výpis je neúplný.' in result['answer']
+    assert 'Výpis je neúplný.' not in result['answer']
+    assert result['complete'] is False and result['answer_coverage']['status'] == 'supported'
     assert 'unsupported_person_work_type_year_binding' in result['attempts'][0]['issues']
 
 

tools/answer_evaluation/answer_coverage.py

@@ -0,0 +1,77 @@
+"""Separate supported answer scope from the search's technical selection limits.
+
+Supported means the delivered facts/excerpts have evidence; it never promises
+exhaustive retrieval, a complete source document or absence of other records.
+"""
+import re
+
+from .context_policy import normalized
+
+
+def response_scope(question: str, *, content: bool = False) -> str:
+    query = normalized(re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', question))
+    # Record lists differ from the numbered assignment of one selected work.
+    records = r'(?:prac\w*|projekt\w*|student\w*|osob\w*|autor\w*|tag\w*|kategori\w*|zdroj\w*)'
+    modifiers = r'(?:(?:vsetk\w*|diplomov\w*|bakalarsk\w*|zaverecn\w*|timov\w*|vyskumn\w*|a|aj)\s+)*'
+    if re.search(r'\b(?:kolko|pocet|spocitaj)\b', query):
+        return 'count'
+    if re.search(rf'\b(?:vsetk\w*|upln\w*|kompletn\w*)\s+{modifiers}{records}\b'
+                 rf'|\b(?:ake|ktore|vypis|zoznam|prehlad)\s+{modifiers}{records}\b', query):
+        return 'list'
+    if content:
+        if re.search(r'\b(?:zadani\w*|ciel\w*)\b', query) and not re.search(r'\b(?:strucn\w*|zhrn\w*)\b', query):
+            return 'assignment'
+        if re.search(r'\b(?:comu|o com|opis\w*|popis\w*|obsah\w*|zhrn\w*|strucn\w*)\b', query):
+            return 'overview'
+    return 'fact'
+
+
+def answer_coverage(answer: dict, context: dict) -> dict:
+    """Describe only the scope of an already validated RAG selection."""
+    requirements = context.get('requirements', {})
+    scope = response_scope(context.get('question', ''), content=requirements.get('content', False))
+    reasons = list(context.get('truncation_reasons') or [])
+    selected = {row.get('source_id') for row in answer.get('evidence_quotes', [])}
+    cited = set(answer.get('source_urls', []))
+    clipped = any('[OBSAH ZDROJA OREZANÝ]' in row.get('body', '')
+                  for row in context.get('source_blocks', [])
+                  if row.get('source_id') in selected or (not selected and
+                     (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
+    limited = context.get('complete') is not True or context.get('truncated') is True
+    limitations = [*reasons]
+    if clipped:
+        limitations.append('cited_source_text_clipped')
+    warnings = []
+    status = answer.get('status')
+    if status == 'ok':
+        coverage = 'supported'
+        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:
+            coverage = 'unknown'; warnings.append('count_not_established')
+        elif scope == 'assignment' and clipped:
+            coverage = 'partial'; warnings.append('assignment_text_clipped')
+        elif scope == 'assignment' and limited and re.search(r'\b(?:vsetk\w*|upln\w*|kompletn\w*|cele)\b', normalized(context.get('question', ''))):
+            coverage = 'partial'; warnings.append('assignment_completeness_unknown')
+    elif status == 'no_answer':
+        coverage = 'not_found'
+        if clipped or context_limited:
+            warnings.append('missing_information_with_clipped_context')
+    else:
+        coverage = 'unavailable'
+    return {'scope': scope, 'status': coverage, 'warnings': warnings,
+            'limitations': list(dict.fromkeys(limitations))}
+
+
+def coverage_notices(coverage: dict) -> list[str]:
+    messages = {
+        'list_may_be_incomplete': 'Uvedený zoznam nemusí obsahovať všetky záznamy.',
+        'assignment_text_clipped': 'Časť podkladov k zadaniu bola skrátená; uvedené body nemusia tvoriť celé zadanie.',
+        'assignment_completeness_unknown': 'Z dostupných podkladov nemožno zaručiť všetky body zadania.',
+        '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']]

tools/answer_evaluation/rag_evidence.py

@@ -313,6 +313,8 @@ def rag_evidence_context(question: str, payload: dict) -> dict:
                     'start_year': start_year,
                     'response_mode': 'claims' if academic or start_year else 'quotes'}
     if content:
+        from .answer_coverage import response_scope
+        requirements['content_style'] = response_scope(question, content=True)
         requirements['response_mode'] = 'mixed' if any(requirements[key] for key in ['title', 'year', 'start_year']) else 'quotes'
     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]

tools/answer_evaluation/verified_rag.py

@@ -10,6 +10,7 @@ from pydantic import BaseModel, ConfigDict, Field
 from .answer_format import Claim
 from .context_policy import answer_context_issues, citation_check, allowed_urls, normalized, text_urls
 from .rag_evidence import WORK_LABELS, descriptive_text, rag_evidence_context
+from .answer_coverage import answer_coverage, coverage_notices
 
 
 class RagClaim(Claim):
@@ -50,6 +51,10 @@ Vyber opisujúce evidence_quotes z content_passages zodpovedajúcej osoby, typu
 Samotný názov, rok, meno vedúceho ani nadpis zadania nenahrádzajú opis.
 Pri response_mode=mixed uveď požadované claims aj opisujúce evidence_quotes.
 Použi iba dodané content_passages; opis inej práce alebo iného roku nie je odpoveď.
+Pri requirements.content_style=overview vyber najviac DVA krátke doslovné výňatky
+na požadovanú osobu, najviac 600 znakov a dve odrážky/body na osobu. Odpovedaj stručne na obsah;
+neprepisuj celé zadanie a históriu stretnutí. Pri content_style=assignment zachovaj
+požadované body zadania; tento stručný limit sa na zadanie nevzťahuje.
 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.
@@ -218,6 +223,49 @@ def rag_answer_issues(answer: dict, context: dict) -> list[str]:
     return list(dict.fromkeys(issues))
 
 
+def overview_excerpts(answer: dict, context: dict) -> list[str]:
+    """Shorten validated quotes by selection, never by factual paraphrase.
+
+    Presentation length is not a reason to reject otherwise supported evidence
+    or to spend another model request rewriting the same facts.
+    """
+    selected = []; counts = {}; sizes = {}; seen = set()
+    for quote in answer.get('evidence_quotes', []):
+        sid = quote['source_id']
+        owner = next((p.get('person') for p in context.get('content_passages', [])
+                      if p['source_id'] == sid and p.get('person')), sid)
+        identities = [row.get(key) for row in context.get('facts', []) if row.get('source_id') == sid
+                      for key in ('person', 'title', 'document_author')]
+        blocks = re.split(r'\n\s*\n|(?=^[ \t]*(?:\d+[.)]|[-+*])\s+)', quote['quote'], flags=re.M)
+        for block in blocks:
+            if counts.get(owner, 0) >= 2:
+                break
+            block = block.strip()
+            if not descriptive_text(block, identities):
+                continue
+            # Prefer complete sentences when a prose paragraph is very long.
+            units = re.split(r'(?<=[.!?])\s+(?=[A-ZÁČĎÉÍĽŇÓŠŤÚÝŽ])', block) if len(block) > 600 else [block]
+            for unit in units:
+                remaining = 600 - sizes.get(owner, 0)
+                if counts.get(owner, 0) >= 2 or remaining < 30:
+                    break
+                unit = unit.strip()
+                if not descriptive_text(unit, identities) or (owner, unit) in seen:
+                    continue
+                if len(unit) > remaining:
+                    # An explicit ellipsis makes a shortened literal prefix
+                    # visible; never cut a word/URL and pretend it is complete.
+                    prefix = unit[:remaining - 2].rsplit(' ', 1)[0].rstrip()
+                    if not prefix or ' ' not in prefix or not descriptive_text(prefix, identities):
+                        continue
+                    unit = prefix + ' …'
+                selected.append(unit); seen.add((owner, unit))
+                counts[owner] = counts.get(owner, 0) + 1
+                sizes[owner] = sizes.get(owner, 0) + len(unit)
+    # An indivisible technical string is better retained than silently omitted.
+    return selected or [q['quote'] for q in answer.get('evidence_quotes', [])]
+
+
 def render_rag_answer(answer: dict, context: dict) -> str:
     lines = []
     if answer['status'] == 'ok':
@@ -227,9 +275,18 @@ def render_rag_answer(answer: dict, context: dict) -> str:
             values = [WORK_LABELS[claim['work_type']]] if claim.get('work_type') else []
             values += [f'{label}: {claim[key]}' for key, label in labels.items() if claim.get(key) is not None]
             lines.append(claim['person'] + ' — ' + '; '.join(values) + '.')
-        sources = {row['source_id']: row for row in context['source_blocks']}
-        lines += ['Zo zdroja ' + str(sources[quote['source_id']].get('title') or quote['source_id']) +
-                  ': „' + quote['quote'] + '“' for quote in answer.get('evidence_quotes', [])]
+        quotes = answer.get('evidence_quotes', [])
+        style = context.get('requirements', {}).get('content_style')
+        if quotes and style in {'overview', 'assignment'}:
+            if style == 'overview':
+                lines.append('Stručne z dostupných podkladov:')
+                lines += overview_excerpts(answer, context)
+            else:
+                lines += [quote['quote'] for quote in quotes]
+        else:
+            sources = {row['source_id']: row for row in context['source_blocks']}
+            lines += ['Zo zdroja ' + str(sources[quote['source_id']].get('title') or quote['source_id']) +
+                      ': „' + quote['quote'] + '“' for quote in quotes]
         if answer.get('count') is not None:
             lines.append(f'Počet: {answer["count"]}.')
     elif answer['status'] == 'no_answer':
@@ -238,10 +295,7 @@ def render_rag_answer(answer: dict, context: dict) -> str:
         lines.append('Údaje sú nejednoznačné; požiadavku treba spresniť.')
     else:
         lines.append('Overenú odpoveď sa nepodarilo pripraviť.')
-    if context.get('truncated') or context.get('complete') is False:
-        lines.append('Výpis je neúplný.')
-    elif context.get('complete') is None:
-        lines.append('Úplnosť vyhľadania nemožno zaručiť.')
+    lines.extend(coverage_notices(answer_coverage(answer, context)))
     if answer.get('source_urls'):
         lines.append('Zdroje: ' + ', '.join(answer['source_urls']))
     return '\n'.join(lines)
@@ -254,6 +308,7 @@ def generate_verified_rag(question: str, payload: dict, complete) -> dict:
         empty = {'status': status, 'claims': [], 'evidence_quotes': [], 'source_urls': []}
         return {'status': status, 'reason': payload.get('reason'), 'accepted_for_delivery': status != 'error',
                 'answer': render_rag_answer(empty, context), 'source_urls': [], 'claims': [],
+                'answer_coverage': answer_coverage(empty, context),
                 'attempts': [], 'index_version': context['index_version'], 'complete': context['complete'],
                 'truncated': context['truncated'], 'truncation_reasons': context['truncation_reasons']}
     sent = {key: context[key] for key in ['status', 'reason', 'complete', 'truncated',
@@ -287,6 +342,7 @@ def generate_verified_rag(question: str, payload: dict, complete) -> dict:
             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'],
+                    'answer_coverage': answer_coverage(answer, context),
                     'attempts': attempts, 'index_version': context['index_version'],
                     'complete': context['complete'], 'truncated': context['truncated'],
                     'truncation_reasons': context['truncation_reasons']}
@@ -299,7 +355,9 @@ def generate_verified_rag(question: str, payload: dict, complete) -> dict:
                     'Nepotrebné citáty a metadáta vynechaj.' if context['requirements']['response_mode'] == 'quotes' else '')
                  + (' Opis práce musí obsahovať doslovné evidence_quotes z content_passages. '
                     'Samotný názov a rok nestačia; pri mixed zachovaj aj požadované claims.'
-                    if context['requirements'].get('content') else '')}]
+                    if context['requirements'].get('content') else '')
+                 + (' Pri overview vyber najviac dva krátke výňatky, najviac 600 znakov a dva body na požadovanú osobu.'
+                    if context['requirements'].get('content_style') == 'overview' else '')}]
     return {'status': 'error', 'reason': 'answer_rejected', 'accepted_for_delivery': False,
             'answer': 'Odpoveď neprešla kontrolou podkladov; neoverené údaje sa nezobrazia.',
             'attempts': attempts, 'index_version': context['index_version']}

tools/rag/rag_utils.py

@@ -29,9 +29,11 @@ NO_ANSWER_TEXT = (
 
 
 RAG_INSTRUCTIONS = [
-    'Rešpektuj complete, truncated, truncation_reasons a index_version. Pri truncated=true alebo '
-    'complete=false v odpovedi jasne uveď, že výpis je neúplný. Pri complete=null nemožno zaručiť '
-    'úplnosť; neoznačuj výber za všetky práce alebo úplný zoznam. Cituj iba URL zo sources.',
+    'Rešpektuj complete, truncated, truncation_reasons a index_version ako rozsah vyhľadania. '
+    'Výber najlepších kandidátov a zdrojov sám osebe nezneplatňuje doložený fakt ani opis. '
+    'Pri požiadavke na zoznam upozorni, ak nemožno zaručiť všetky záznamy. Pri orezanom '
+    'texte a chýbajúcej informácii pomenuj konkrétne obmedzenie. Nikdy neoznačuj neúplný '
+    'alebo neistý výber za úplný zoznam. Cituj iba URL zo sources.',
     (
         "Odpovedaj výhradne podľa informácií vrátených "
         "v poli context a sources zo ZP Wiki."