Even after format_quotes_for_prompt was fixed to expose the fully-qualified programme name + page in each chunk, Qwen continued to hallucinate familiar source labels (typically "CDU Wahlprogramm NRW 2022") for parties whose actual MV/BE chunks were thematically off-topic for the Antrag at hand. The model preferred its training prior over the prompt context. Smoke test: MV Drucksache 8/6390 (CDU "Krisenmechanismus Kraftstoffpreise"). The CDU MV chunks the embedder retrieved were about Senioren and Aussenwirtschaft — not about energy prices — so qwen pulled what it knew about CDU NRW transport policy and wrote that as the source, even though the prompt listed only "CDU Mecklenburg-Vorpommern Wahlprogramm 2021" chunks. The new explicit ZITATEREGEL block in the user prompt forbids cross-Bundesland citations and instructs the model to leave zitate empty rather than fabricate when no listed chunk fits. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| kontext | ||
| routers | ||
| static/referenzen | ||
| templates | ||
| __init__.py | ||
| analyzer.py | ||
| bundeslaender.py | ||
| config.py | ||
| database.py | ||
| embeddings.py | ||
| main.py | ||
| models.py | ||
| parlamente.py | ||
| report.py | ||
| wahlprogramme.py | ||