Oprava grafových podmienok

README.md

@@ -1523,6 +1523,16 @@ sa nepoužije ako dôkaz dátumu jej začatia či obhajoby. Chybové odpovede od
 nedostupné vyhľadávanie, nedostupný model, vyťaženie a zamietnutý neoverený text;
 raw výstupy odmietnutého modelu sa štandardnému klientovi neposkytujú.
 
+Grafový filter tagu sa uplatňuje aj pri otázkach na neurčených študentov a
+konkrétny typ/rok práce: všetky podmienky musí spĺňať tá istá osoba. Spojenie
+„tag RAG a zároveň diplomový projekt“ nepatrí do názvu tagu. Úvodzovky umožňujú
+zachovať viacslovný tag aj so spojkou „a“. Pri názve práce konkrétnej osoby sa
+zachovajú aj záznamy bez názvu; odpoveď uvedie známy typ a rok a chýbajúci názov.
+Spoločné tagy sa zobrazujú súhrnne, s dôkazmi pre každú požadovanú osobu.
+Rozpoznané podmienky na počet vrátane slovných čísloviek („aspoň tri“) vyžadujú
+spresnenie; prahové filtrovanie zatiaľ nie je implementované. Predikcia dokončenia
+práce sa výslovne odlišuje od vyhľadania práce evidovanej v určitom roku.
+
 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

test/test_graph_query_regressions.py

@@ -0,0 +1,191 @@
+"""Regressions for actual Claude questions and adversarial same-person joins."""
+import copy
+import json
+from pathlib import Path
+
+import pytest
+
+from tools.answer_evaluation.checked_generation import generate_checked_answer
+from tools.answer_evaluation.context_policy import answer_context_issues
+from tools.answer_evaluation.graph_context import graph_context
+from tools.answer_evaluation.render_answer import render_graph_answer
+from tools.answers.answer_delivery import deliver_answer
+from tools.graphrag import graph_search as gs
+from tools.graphrag.tests.test_graph_search import FakeSession
+from tools.retrieval.question_constraints import unsupported_constraint
+
+
+PEOPLE=['Ján Pták','Ján Malinovský','Valerii Kutsenko','Jakub Schwarc']
+PROMPT=Path('tools/answer_evaluation/graph_answer_prompt.txt').read_text()
+
+
+def person(index,name):
+    return {'kind':'Person','name':name,'element_id':f'p{index}',
+            'source_document':f'pages/students/2022/p{index}/README.md'}
+
+
+def source(index):
+    return {'document_path':f'pages/students/2022/p{index}/README.md',
+            'source_url':f'https://example.test/p{index}'}
+
+
+@pytest.fixture
+def tagged_projects():
+    catalog=[person(i,name) for i,name in enumerate(PEOPLE,1)]
+    catalog.append({'kind':'Keyword','name':'RAG','id':'rag','element_id':'krag'})
+    projects=[{'person':name,'person_element_id':f'p{i}','year':2026,
+               'work_type':'diploma_project','work_id':f'work:p{i}',
+               'heading':'Diplomový projekt 2026','title':f'Téma projektu {name}',**source(i)}
+              for i,name in enumerate(PEOPLE,1)]
+    tags=[{'person':name,'person_element_id':f'p{i}','keyword':'RAG',
+           'keyword_id':'rag','keyword_element_id':'krag',**source(i)}
+          for i,name in enumerate(PEOPLE[:3],1)]
+    # A fifth person has the tag, but no work in the requested year/type.
+    tags.append({'person':'Iná Osoba','person_element_id':'p5','keyword':'RAG',
+                 'keyword_id':'rag','keyword_element_id':'krag',**source(5)})
+    return FakeSession({gs.CATALOG_QUERY:catalog,gs.WORK_FACTS_QUERY:projects,
+                        gs.PERSON_KEYWORD_FACTS_QUERY:tags})
+
+
+@pytest.mark.parametrize('question',[
+    'Ktorí študenti majú tag RAG a zároveň diplomový projekt z roku 2026?',
+    'Ktorí študenti s tagom RAG majú diplomový projekt v roku 2026?',
+    'Ktorí študenti majú tag RAG a diplomový projekt z roku 2026?',
+    'Ktorí študenti majú tag „RAG“ a zároveň diplomový projekt z roku 2026?',
+    'Ktorí študenti s tagom RAG a zároveň diplomovým projektom z roku 2026?',
+    'Ktorí študenti s tagom RAG a diplomovým projektom z roku 2026?',
+])
+def test_tag_and_project_are_witnessed_by_the_same_person(tagged_projects,question):
+    r=gs.search_graph_with_session(tagged_projects,question,limit=50)
+    assert r['status']=='ok' and r['constraints']['keyword_phrase']=='rag'
+    assert {f['person'] for f in r['facts']}==set(PEOPLE[:3])
+    work=[params for q,params in tagged_projects.calls if q==gs.WORK_FACTS_QUERY][0]
+    tags=[params for q,params in tagged_projects.calls if q==gs.PERSON_KEYWORD_FACTS_QUERY][0]
+    assert work['years']==[2026] and work['work_types']==['diploma_project']
+    assert tags['keyword_ids']==['krag']
+    context=graph_context(r)
+    assert {c['person'] for c in context['answer_candidates']}==set(PEOPLE[:3])
+    assert all(c['work_type']=='diploma_project' for c in context['answer_candidates'])
+
+
+def test_diploma_thesis_cannot_be_replaced_by_a_project(tagged_projects):
+    r=gs.search_graph_with_session(tagged_projects,
+        'Ktorí študenti majú tag RAG a zároveň diplomovú prácu z roku 2026?',limit=50)
+    assert r['status']=='no_answer' and r['reason']=='no_matching_facts'
+    assert not r['facts'] and not r['constraints']['unresolved_keyword']
+
+
+def test_unknown_tag_cannot_trigger_a_global_project_listing(tagged_projects):
+    r=gs.search_graph_with_session(tagged_projects,
+        'Ktorí študenti majú tag XNEZNAMY a zároveň diplomový projekt z roku 2026?')
+    assert r['status']=='no_answer' and r['reason']=='unknown_keyword' and not r['facts']
+
+
+@pytest.mark.parametrize('question,expected',[
+    ('Ľudia s tagom strojové učenie a zároveň diplomovou prácou.','strojove ucenie'),
+    ('Ľudia s tagom „výskum a aplikácie“ a zároveň diplomovým projektom.','vyskum a aplikacie'),
+    ('Ľudia s tagom RAG z roku 2026.','rag'),
+])
+def test_multiword_tag_and_condition_boundary(question,expected):
+    assert gs.explicit_keyword_phrase(question)==expected
+
+
+@pytest.mark.parametrize('question,code',[
+    ('Ktorí študenti majú aspoň tri práce?','quantity_condition'),
+    ('Ktorí študenti majú aspoň 3 práce?','quantity_condition'),
+    ('Kto má najmenej tri práce?','quantity_condition'),
+    ('Študenti s najviac dvoma projektmi?','quantity_condition'),
+    ('Študenti s minimálne piatimi prácami?','quantity_condition'),
+    ('ktorý študenti týmto tempom ukončia svoje diplomové práce v roku 2027?','completion_prediction'),
+    ('Ktorí študenti ukončia diplomové práce v roku 2027?','completion_prediction'),
+    ('Odhadni termín dokončenia diplomovej práce Jána Holpa.','completion_prediction'),
+    ('Kedy Ján Holp ukončil diplomovú prácu?','work_event_date'),
+    ('Aké metódy predikcie počasia používa diplomová práca?',None),
+    ('Aký je názov práce „Kto dokončí prácu v roku 2027“?',None),
+])
+def test_unsupported_condition_wording(question,code):
+    assert unsupported_constraint(question)==code
+
+
+@pytest.mark.parametrize('engine',['rag','graphrag'])
+@pytest.mark.parametrize('question,word',[
+    ('Ktorí študenti majú aspoň tri práce?','počet'),
+    ('Ktorí študenti týmto tempom ukončia diplomové práce v roku 2027?','predpovedať'),
+])
+def test_unsupported_condition_never_calls_a_model(engine,question,word):
+    r=deliver_answer(question,engine=engine,complete=lambda _:pytest.fail('No model for unsupported condition'))
+    assert r['status']=='ambiguous' and r['accepted_for_delivery'] and word in r['answer']
+    assert not r['claims'] and not r['source_urls']
+
+
+def selected(context):
+    fields=['person','record_id','title','year','work_type','work_label','keyword','category','document_author']
+    rows=context.get('required_academic_records') or context['facts']
+    claims=[{k:r[k] for k in fields if r.get(k) is not None} for r in rows]
+    return {'status':'ok','answer':'','claims':claims,
+            'source_urls':[s['source_url'] for s in context['sources']]}
+
+
+def generated(answer,context):
+    return generate_checked_answer(context.get('question',''),context,PROMPT,
+        lambda _: {'choices':[{'message':{'content':json.dumps(answer)}}]})
+
+
+def test_titleless_records_survive_and_are_delivered_without_invented_titles():
+    catalog=[person(1,'Ján Holp')]
+    records=[{'person':'Ján Holp','person_element_id':'p1','work_id':f'project-{i}',
+              'title':None,'heading':label,'work_type':'diploma_project','year':2020,**source(1)}
+             for i,label in enumerate(['Diplomový projekt 2020','Diplomový projekt 2 2020'])]
+    s=FakeSession({gs.CATALOG_QUERY:catalog,gs.WORK_FACTS_QUERY:records})
+    raw=gs.search_graph_with_session(s,'Aký názov má diplomový projekt Jána Holpa z roku 2020?')
+    assert raw['status']=='ok' and len(raw['facts'])==2
+    context=graph_context(raw);a=selected(context)
+    result=generated(a,context)
+    assert result['accepted_for_delivery']
+    assert result['answer']['answer'].count('názov v zdroji nie je uvedený')==2
+    assert len(result['answer']['claims'])==2 and not any(c.get('title') for c in result['answer']['claims'])
+    bad=copy.deepcopy(a);bad['claims'][0]['title']='Vymyslený názov'
+    assert 'unsupported_academic_binding' in answer_context_issues(bad,context,check_text=False)
+    bad=copy.deepcopy(a);bad['claims'].pop()
+    assert 'missing_academic_record_details' in answer_context_issues(bad,context,check_text=False)
+
+
+def shared_context(three=False):
+    people=PEOPLE[:3] if three else PEOPLE[:2]
+    raw={'status':'ok','reason':None,'complete':True,'truncated':False,'aggregation':None,
+         'query':'Aké spoločné kľúčové slová majú '+' a '.join(people)+'?',
+         'constraints':gs.extract_query_constraints('spoločné kľúčové slová'),
+         'resolved_entities':[person(i,p) for i,p in enumerate(people,1)],
+         'sources':[source(i) for i in range(1,len(people)+1)],'facts':[]}
+    for keyword,left,right in [('NLP',1,2),('RAG',1,2)]+([('NLP',1,3)] if three else []):
+        raw['facts'].append({'fact_type':'shared_keyword','keyword':keyword,
+            'left_person':people[left-1],'right_person':people[right-1],
+            'left_document_path':source(left)['document_path'],'right_document_path':source(right)['document_path'],
+            'left_source_url':source(left)['source_url'],'right_source_url':source(right)['source_url']})
+    return graph_context(raw)
+
+
+def test_shared_tags_are_summarized_once_and_require_all_peoples_evidence():
+    context=shared_context();a=selected(context)
+    result=generated(a,context)
+    assert result['accepted_for_delivery']
+    text=result['answer']['answer']
+    assert 'Spoločné kľúčové slová: **NLP**, **RAG**.' in text
+    assert text.count('NLP')==1 and text.count('RAG')==1 and 'Pri osobe' not in text
+    bad=copy.deepcopy(a);bad['claims'].pop()
+    assert 'missing_shared_property_evidence' in answer_context_issues(bad,context,check_text=False)
+
+
+def test_pairwise_shared_tag_does_not_mean_shared_by_everyone():
+    context=shared_context(three=True)
+    assert {r['value'] for r in context['shared_properties']}=={'NLP'}
+    text=render_graph_answer(selected(context),context)['answer']
+    assert 'Spoločné kľúčové slová: **NLP**.' in text and '**RAG**' not in text
+
+
+def test_incomplete_shared_tags_cannot_claim_exhaustive_results():
+    context=shared_context();context.update(complete=False,truncated=True,truncation_reasons=['fact_limit'])
+    result=generated(selected(context),context)
+    assert result['accepted_for_delivery']
+    assert 'Doložené spoločné kľúčové slová' in result['answer']['answer']
+    assert 'Výpis je neúplný.' in result['answer']['answer']

tools/answer_evaluation/context_policy.py

@@ -76,6 +76,8 @@ def joined_candidates(facts: list[dict], intent: dict) -> list[dict]:
                 continue
             fields = ['person', 'title', 'year'] if needs_title else ['person', 'year']
             candidate={k:record[k] for k in fields if record.get(k) is not None}
+            if record.get('work_type'):
+                candidate['work_type']=record['work_type']
             candidate['keyword']=tag['keyword']
             candidate['evidence_documents']=sorted({p for p in [record.get('document_path'),tag.get('document_path')] if p})
             candidate['source_urls']=sorted({u for u in [record.get('source_url'),tag.get('source_url')] if u})
@@ -220,11 +222,16 @@ def answer_context_issues(answer: dict, context: dict, *, check_text: bool = Tru
                        for claim in claims):
                 issues.append('missing_academic_record_details')
                 break
+    for shared in context.get('shared_properties',[]):
+        for person in shared['people']:
+            if not any(normalized(c.get('person'))==normalized(person) and
+                       normalized(c.get(shared['field']))==normalized(shared['value']) for c in claims):
+                issues.append('missing_shared_property_evidence')
     candidates=context.get('answer_candidates',[])
     if context.get('status')=='ok' and candidates:
         if answer.get('status')=='no_answer':
             issues.append('rejected_supported_person_join')
-        elif not all(any(all(normalized(claim.get(k))==normalized(candidate[k]) for k in ['person','title','year','keyword'] if k in candidate)
+        elif not all(any(all(normalized(claim.get(k))==normalized(candidate[k]) for k in ['person','title','year','keyword','work_type'] if k in candidate)
                          for claim in answer.get('claims',[])) for candidate in candidates):
             issues.append('split_or_missing_person_join')
     if context.get('status') in {'no_answer','ambiguous','error'} and answer.get('status')=='ok':

tools/answer_evaluation/graph_answer_prompt.txt

@@ -6,6 +6,7 @@ Title je názov akademickej práce/projektu, nikdy názov profilovej stránky al
 Všetky podmienky jednej claim patria tej istej osobe. Title a year patria rovnakému akademickému záznamu. Category a document_author musia byť doložené tým istým document_path.
 Tag patrí osobe cez jej profil aj dcérske dokumenty. Pre kombináciu názvu/roku a tagu nevyžaduj rovnaký dokument. Answer_candidates sú už doložené spojenia cez tú istú osobu; zachovaj ich person, year, keyword a prípadne title spolu v jednej claim. Neodmietaj toto spojenie len pre odlišné document_path.
 Claims musia pokryť všetky relevantné výsledky dotazu a požadované väzby. Pri porovnaní ľudí uveď každého v samostatnej claim. Pri výbere kandidátov vynechaj nevyhovujúce osoby. Pri názve práce uveď aj rok, ak je vo faktoch známy.
+Ak sú dodané shared_properties, každú spoločnú hodnotu dolož claim pre každú uvedenú osobu. Text spoločného výsledku zostaví aplikácia. Pri answer_candidates zachovaj aj work_type, ak je dodaný: diplomový projekt a diplomová práca sú odlišné typy. Required_academic_records môžu zahŕňať aj záznamy bez názvu pri otázke na názov; zachovaj ich známy typ, rok a identitu, ale názov nevymýšľaj.
 Academic_year_comparison opisuje spoločné roky všetkých požadovaných osôb, nie zjednotenie ich rokov. Pri dodaných required_academic_records zachovaj aj typy, názvy a identity prác z týchto rokov. Nepridávaj práce z iných rokov.
 Source_urls musí obsahovať konkrétne relevantné source_url zo sources pre všetky tvrdenia. Nevymýšľaj ani neupravuj URL. Ak názov/rok dokladá profil a tag dcérsky dokument, cituj oba potrebné zdroje. Nepoužívaj URL iných osôb ako náhradnú citáciu.
 Count môže byť číslo iba pri aggregation.complete=true a musí sa rovnať aggregation.count. Inak count=null, aj pri no_answer. Počet nie je počet zobrazených zdrojov. Presný agregovaný počet môže byť úplný aj pri orezanom výpise dôkazov.

tools/answer_evaluation/graph_context.py

@@ -10,7 +10,10 @@ def graph_context(result: dict) -> dict:
     """
     facts = []; intent = result['constraints']
     shared_year_query = intent.get('mentions_shared') and intent.get('mentions_academic_year')
-    work_listing = intent.get('mentions_work_listing', False) or shared_year_query
+    work_listing = (intent.get('mentions_work_listing', False) or shared_year_query or
+                    (intent.get('mentions_title') and any(
+                        row.get('fact_type') in {'person_thesis','person_work'} and not row.get('title')
+                        for row in result['facts'])))
     academic_records = []
     for row in result['facts']:
         kind = row.get('fact_type', '')
@@ -33,6 +36,10 @@ def graph_context(result: dict) -> dict:
         if intent['mentions_start_year']:
             allowed.add('start_year')
         fact = {k: v for k, v in row.items() if k in allowed and v is not None}
+        if (intent.get('thesis_types') or intent.get('work_types')) and kind in {'person_thesis','person_work'}:
+            work_type=row.get('thesis_type') or row.get('work_type')
+            if work_type:
+                fact['work_type']=work_type
         if work_listing and kind in {'person_thesis', 'person_work'}:
             # Stable source record identity distinguishes two untitled projects
             # by the same person in the same year. Never use Neo4j element IDs.
@@ -57,6 +64,14 @@ def graph_context(result: dict) -> dict:
     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')))
+    shared_values={}
+    for fact in facts:
+        for field in ['keyword','category','document_author']:
+            if intent.get('mentions_shared') and fact.get(field) is not None:
+                shared_values.setdefault((field,fact[field]),set()).add(fact['person'])
+    people=context['requested_people']
+    context['shared_properties']=[{'field':field,'value':value,'people':people}
+        for (field,value),witnesses in shared_values.items() if len(people)>=2 and set(people)<=witnesses]
     comparison = result.get('academic_year_comparison') or next((row for row in result['facts']
         if row.get('fact_type') == 'academic_year_comparison'), None)
     if comparison is not None:

tools/answer_evaluation/render_answer.py

@@ -10,6 +10,16 @@ def render_graph_answer(answer: dict, context: dict) -> dict:
     out = deepcopy(answer)
     lines = []
     if out['status'] == 'ok':
+        shared_fields=set()
+        labels={'keyword':'kľúčové slová','category':'kategórie','document_author':'autori dokumentov'}
+        for field,label in labels.items():
+            values=sorted({row['value'] for row in context.get('shared_properties',[]) if row['field']==field
+                and all(any(c.get('person')==person and c.get(field)==row['value'] for c in out['claims'])
+                        for person in row['people'])})
+            if values:
+                shared_fields.add(field)
+                prefix='Spoločné ' if context.get('complete') is True and not context.get('truncated') else 'Doložené spoločné '
+                lines.append(prefix+label+': '+', '.join('**'+v+'**' for v in values)+'.')
         comparison = context.get('academic_year_comparison')
         if comparison:
             people = comparison.get('people', [])
@@ -27,7 +37,9 @@ def render_graph_answer(answer: dict, context: dict) -> dict:
             if claim.get('record_id'):
                 records.setdefault(claim['person'], []).append(claim)
             else:
-                lines.append(fact_sentences(claim, context.get('question', '')))
+                shown={key:value for key,value in claim.items() if key not in shared_fields}
+                if any(key!='person' and value is not None for key,value in shown.items()):
+                    lines.append(fact_sentences(shown, context.get('question', '')))
         index = 0
         for person, claims in records.items():
             lines.extend([f'{person} má v dostupných podkladoch tieto práce a projekty:', ''])

tools/graphrag/graph_search.py

@@ -725,6 +725,13 @@ def has_stem(
 
 def explicit_keyword_phrase(query: str) -> str | None:
     """Identify a requested tag without treating 'ZP Wiki' as the tag ZP."""
+    quoted = re.search(
+        r'\b(?:kľúčové\s+slovo|kľúčovým\s+slovom|kľúčového\s+slova|'
+        r'klucove\s+slovo|klucovym\s+slovom|klucoveho\s+slova|tagom|tagu|tag)'
+        r'\s+[„“"«]([^„“”"»]+)[“”"»]', query, flags=re.I,
+    )
+    if quoted:
+        return normalize_for_match(quoted[1])
     match = re.search(
         r"\b(?:klucove\s+slovo|klucovym\s+slovom|klucoveho\s+slova|"
         r"tagom|tagu|tag)\s+(.+)",
@@ -733,7 +740,9 @@ def explicit_keyword_phrase(query: str) -> str | None:
     if not match:
         return None
     phrase = re.split(
-        r"\s+(?:pri\s+osob\w*|u\s+osob\w*|a\s+(?:jej|jeho|ich|rok|pomenuj|uved\w*)|ak\s+|uved\w*\s+|"
+        r"\s+(?:pri\s+osob\w*|u\s+osob\w*|a\s+(?:zaroven|sucasne|diplomov\w*|bakalarsk\w*|"
+        r"dizertacn\w*|timov\w*|vyskumn\w*|projekt\w*|prac\w*|jej|jeho|ich|rok|pomenuj|uved\w*)|"
+        r"zaroven\b|sucasne\b|(?:v|z|pre)\s+rok\w*\b|ak\s+|uved\w*\s+|"
         r"pomenuj\s+|maju\s+|ma\s+)",
         match.group(1),
         maxsplit=1,
@@ -1241,7 +1250,7 @@ def requested_name_phrases(query: str) -> list[str]:
     )
     if role:
         tail = re.split(
-            r'\s+(?:s(?:o)?\s+(?:nazv\w*|nazov\w*|rok\w*)|splna\w*|ma|maju|mal|mali|je|su|'
+            r'(?:^|\s+)(?:s(?:o)?\s+(?:nazv\w*|nazov\w*|rok\w*)|splna\w*|ma|maju|mal|mali|je|su|'
             r'ktory\w*|ktora\w*|pri\s+ktorych|v\s+roku|uved\w*|prepojen\w*|spojen\w*|'
             r'vyber|najdi|zisti|cez|pomenuj|uved\w*|vypis|ak)\b', role.group(1), maxsplit=1,
         )[0]
@@ -1263,6 +1272,7 @@ def requested_name_phrases(query: str) -> list[str]:
     return list(dict.fromkeys(
         candidate.strip() for candidate in candidates
         if 1 <= len(candidate.split()) <= 5 and candidate.split()[0] not in generic
+        and not re.match(r'^(?:diplomov|bakalarsk|dizertacn|timov|vyskumn|zaverecn)\w*\s+(?:prac\w*|projekt\w*)\b',candidate)
     ))
 
 
@@ -2448,7 +2458,7 @@ def search_graph_with_session(
     )
     keyword_specific = bool(
         not shared_keyword and (
-            (keyword_ids and not (thesis_specific or work_specific))
+            keyword_ids
             or (person_ids and constraints["mentions_keyword"])
         )
     )
@@ -2488,7 +2498,9 @@ def search_graph_with_session(
     if constraints["mentions_count"] and not (thesis_specific or work_specific):
         blocked_reasons.append("unsupported_count_domain")
     # A quoted unknown work title is a failed exact lookup, not a global listing.
-    quoted_titles = re.findall(r'[„“"«]([^„“”"»]+)[“”"»]', clean_query)
+    quoted_titles = [match[1] for match in re.finditer(r'[„“"«]([^„“”"»]+)[“”"»]', clean_query)
+                     if not re.search(r'\b(?:tag|tagom|tagu|klucov\w*\s+slov\w*)$',
+                                      normalize_for_match(clean_query[:match.start()]))]
     if quoted_titles and (thesis_specific or work_specific) and not thesis_ids and not work_ids:
         blocked_reasons.append("unknown_title")
         constraints["unresolved_title"] = True
@@ -2509,7 +2521,7 @@ def search_graph_with_session(
     work_rows = fetch(work_specific, WORK_FACTS_QUERY,
                       work_ids=work_ids, years=constraints["years"],
                       work_types=constraints["work_types"])
-    if title_specific and not constraints["mentions_work_listing"]:
+    if title_specific and not constraints["mentions_work_listing"] and not person_ids:
         thesis_rows = [row for row in thesis_rows if row.get("title")]
         work_rows = [row for row in work_rows if row.get("title")]
     start_rows = fetch(constraints["mentions_start_year"], START_YEAR_FACTS_QUERY,

tools/graphrag/tests/test_evaluation_wide.py

@@ -80,13 +80,17 @@ def test_last_leading_title_wins_over_introductory_description():
     assert extract_academics(doc)[0]['title'] == 'Generovanie jazyka s podporou vyhľadávania'
 
 
-def test_title_lookup_does_not_claim_untitled_work_as_an_answer():
+def test_title_lookup_preserves_known_record_without_claiming_a_title():
     fact = {'person_element_id': 'p2', 'person': 'Jakub Maruniak', 'thesis_element_id': 'th1',
             'year': 2024, 'title': None, 'thesis_type': 'diploma_thesis'}
     session = FakeSession({gs.CATALOG_QUERY: sample_catalog(), gs.THESIS_FACTS_QUERY: [fact]})
     result = gs.search_graph_with_session(session, 'Aký názov práce má osoba Jakub Maruniak v roku 2024?')
-    assert result['status'] == 'no_answer'
-    assert result['facts'] == []
+    # A missing title must not erase the independently known work/year. The
+    # delivery regression checks the explicit "title missing" presentation.
+    assert result['status'] == 'ok'
+    assert len(result['facts']) == 1
+    assert result['facts'][0]['title'] is None
+    assert result['facts'][0]['year'] == 2024
 
 
 def test_explicit_name_precedes_a_different_topic_and_link_title_is_preserved():

tools/retrieval/question_constraints.py

@@ -14,13 +14,23 @@ def unsupported_constraint(question: str) -> str | None:
                    if not unicodedata.combining(c))
     # Do not treat words inside quoted academic titles as query operators.
     text = re.sub(r'[„“"«][^„“”"»]+[“”"»]', ' ', text)
+    academic = re.search(r'\b(?:prac\w*|projekt\w*)\b',text)
+    completion = re.search(r'\b(?:ukonc\w*|dokonc\w*|obhaj\w*|odovzd\w*|stihn\w*)\b',text)
+    if academic and completion and (re.search(r'\b(?:tymto tempom|predpoved\w*|predik\w*|stihne|stihnu)\b',text) or
+                     re.search(r'\b(?:ukonci|ukoncia|dokonci|dokoncia|obhaji|obhaja)\b',text) or
+                     (re.search(r'\bodhad\w*\b',text) and
+                      re.search(r'\b(?:ukonc\w*|dokonc\w*|obhaj\w*)\b',text))):
+        return 'completion_prediction'
     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)):
+        re.search(r'\b(?:obhajob\w*|obhajil\w*|dokonc\w*|ukonc\w*|zacal\w* pracovat)\b',text) and academic):
         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):
+    number = (r'(?:\d+|jeden|jedna|jedno|jednu|jednej|jedneho|dva|dve|dvoch|dvom|dvoma|'
+              r'tri|troch|trom|troma|styri|styroch|styrom|pat\w*|piat\w*|sest\w*|siest\w*|'
+              r'sedem\w*|siedm\w*|osem\w*|osm\w*|devat\w*|deviat\w*|desat\w*|desiat\w*|'
+              r'jedenast\w*|dvanast\w*|trinast\w*|strnast\w*|dvadsat\w*)')
+    if re.search(r'\b(?:aspon|najmenej|najviac|viac ako|menej ako|minimalne|maximalne)\s+'+number+r'\b', 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):
@@ -32,6 +42,8 @@ def unsupported_constraint(question: str) -> str | None:
 
 def constraint_message(code: str) -> str:
     return {
+        'completion_prediction':'Z dostupných poznámok neviem spoľahlivo predpovedať, kto a kedy dokončí prácu. '
+                                'Môžem vyhľadať práce evidované v konkrétnom roku; to nie je odhad ich dokončenia.',
         '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.',