fix: max_tokens extraction CPAM et validation adversariale 1500→3000
Les deux appels tronquaient systématiquement (done_reason=length), causant des JSON invalides et des faux positifs adversariaux. num_predict n'a aucun impact sur VRAM ni sur les réponses courtes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
@@ -91,9 +91,9 @@ def _extraction_pass(
|
||||
)
|
||||
|
||||
logger.debug(" Passe 1 — extraction structurée")
|
||||
result = call_ollama(prompt, temperature=0.0, max_tokens=1500, role="cpam")
|
||||
result = call_ollama(prompt, temperature=0.0, max_tokens=3000, role="cpam")
|
||||
if result is None:
|
||||
result = call_anthropic(prompt, temperature=0.0, max_tokens=1500)
|
||||
result = call_anthropic(prompt, temperature=0.0, max_tokens=3000)
|
||||
if result is not None:
|
||||
logger.info(" Passe 1 OK : %d éléments cliniques extraits",
|
||||
len(result.get("elements_cliniques_pertinents", [])))
|
||||
|
||||
Reference in New Issue
Block a user