Dom/t2a_v2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: max_tokens CPAM 6000→16000 + diagnostic troncature Ollama

Browse Source 
- Argumentation + correction : max_tokens porté à 16000 (num_predict)
- ollama_client : log done_reason=length pour détecter les troncatures serveur
- Résultat live : 1/3 Tier B (dossier 132 passé de C à B, score 5/10)
- Les 2 Tier C restants sont bloqués par hallucination de codes et
  absence de données bio, pas par max_tokens

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
dom
2026-02-20 15:00:08 +01:00
parent 4d49d4e114
commit 8e0ed1220d
2 changed files with 13 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/control/cpam_response.py
View File

@@ -126,12 +126,12 @@ def generate_cpam_response(
prompt, tag_map = _build_cpam_prompt(dossier, controle, sources, extraction)
# 4. Appel LLM — Ollama (rôle cpam) > Haiku fallback
result = call_ollama(prompt, temperature=0.1, max_tokens=6000, role="cpam")
result = call_ollama(prompt, temperature=0.1, max_tokens=16000, role="cpam")
if result is not None:
logger.info(" Contre-argumentation via Ollama")
else:
logger.info(" Ollama indisponible → fallback Anthropic Haiku")
result = call_anthropic(prompt, temperature=0.1, max_tokens=6000)
result = call_anthropic(prompt, temperature=0.1, max_tokens=16000)
if result is not None:
logger.info(" Contre-argumentation via Anthropic Haiku")
@@ -188,9 +188,9 @@ def generate_cpam_response(
validation.get("score_confiance"), len(erreurs_v))
correction_prompt = _build_correction_prompt(prompt, result, validation)
corrected = call_ollama(correction_prompt, temperature=0.0, max_tokens=6000, role="cpam")
corrected = call_ollama(correction_prompt, temperature=0.0, max_tokens=16000, role="cpam")
if corrected is None:
corrected = call_anthropic(correction_prompt, temperature=0.0, max_tokens=6000)
corrected = call_anthropic(correction_prompt, temperature=0.0, max_tokens=16000)
if corrected:
# Re-valider la correction

10
src/medical/ollama_client.py
View File

@@ -189,7 +189,15 @@ def call_ollama(
time.sleep(delay)
continue
response.raise_for_status()
raw = response.json().get("response", "")
resp_data = response.json()
raw = resp_data.get("response", "")
done_reason = resp_data.get("done_reason", "")
eval_count = resp_data.get("eval_count", 0)
if done_reason == "length":
logger.warning("Ollama : réponse tronquée (done_reason=length, %d tokens, %d chars)",
eval_count, len(raw))
else:
logger.debug("Ollama : réponse complète (%d tokens, %d chars)", eval_count, len(raw))
result = parse_json_response(raw)
if result is not None:
return result