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

feat: architecture multi-modèles LLM + quality engine + benchmark

Browse Source 
- Multi-modèles : 4 rôles LLM (coding=gemma3:27b-cloud, cpam=gemma3:27b-cloud,
  validation=deepseek-v3.2:cloud, qc=gemma3:12b) avec get_model(role)
- Prompts externalisés : 7 templates dans src/prompts/templates.py
- Cache Ollama : modèle stocké par entrée (migration auto ancien format)
- call_ollama() : paramètre role= (priorité: model > role > global)
- Quality engine : veto_engine + decision_engine + rules_router (YAML)
- Benchmark qualité : scripts/benchmark_quality.py (A/B, métriques CIM-10)
- Fix biologie : valeurs qualitatives (troponine négative) non filtrées
- Fix CPAM : gemma3:27b-cloud au lieu de deepseek (JSON tronqué par thinking)
- CPAM max_tokens 4000→6000, viewer admin multi-modèles
- Benchmark 10 dossiers : 100% DAS valides, 10/10 CPAM, 243s/dossier

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
dom
2026-02-20 00:21:09 +01:00
parent 5c8c2817ec
commit 909e051cc9
39 changed files with 5092 additions and 574 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
docs/prompts.md Normal file
View File

@@ -0,0 +1,33 @@
# Prompts LLM — Pipeline T2A v2
7 prompts externalisés dans `src/prompts/templates.py`, importables via `from src.prompts import ...`.
| # | Template | Rôle LLM | Modèle par défaut | Temp. | max_tokens | Variables | Appelant |
|---|----------|----------|-------------------|-------|------------|-----------|----------|
| 1 | `CODING_CIM10` | coding | gemma3:27b-cloud | 0.1 | 2500 | texte, type_diag, ctx_str, sources_text | `rag_search._build_prompt()` |
| 2 | `CODING_CCAM` | coding | gemma3:27b-cloud | 0.1 | 2500 | texte, ctx_str, sources_text | `rag_search._build_prompt_ccam()` |
| 3 | `DAS_EXTRACTION` | coding | gemma3:27b-cloud | 0.1 | 2000 | dp_texte, existing_str, ctx_str, text_medical | `rag_search._build_prompt_das_extraction()` |
| 4 | `QC_VALIDATION` | qc | gemma3:12b | 0.1 | 2500 | ctx_str, codes_section | `cim10_extractor._validate_justifications()` |
| 5 | `CPAM_EXTRACTION` | cpam | deepseek-v3.2:cloud | 0.0 | 1500 | dp_str, das_str, tagged_text, titre, arg_ucr, decision_ucr, dp_ucr_line, da_ucr_line | `cpam_response._extraction_pass()` |
| 6 | `CPAM_ARGUMENTATION` | cpam | deepseek-v3.2:cloud | 0.1 | 4000 | dossier_str, asymetrie_str, tagged_str, titre, arg_ucr, decision_ucr, codes_str, definitions_str, sources_text, extraction_str | `cpam_response._build_cpam_prompt()` |
| 7 | `CPAM_ADVERSARIAL` | validation | deepseek-v3.2:cloud | 0.0 | 800 | response_json, factual_section, normes_section, dp_ucr_line, da_ucr_line | `cpam_response._validate_adversarial()` |
## Rôles LLM (config.py)
```python
OLLAMA_MODELS = {
"coding": "gemma3:27b-cloud", # Codage CIM-10/CCAM, extraction DAS
"cpam": "deepseek-v3.2:cloud", # Passe 1 extraction + passe 2 argumentation CPAM
"validation": "deepseek-v3.2:cloud", # Validation adversariale (DOIT différer du cpam en prod)
"qc": "gemma3:12b", # Validation batch justifications, rapide
}
```
Surchargeable par env : `T2A_MODEL_CODING`, `T2A_MODEL_CPAM`, `T2A_MODEL_VALIDATION`, `T2A_MODEL_QC`.
## Priorité de résolution du modèle
`call_ollama(model=, role=)` :
1. `model` explicite (prioritaire)
2. `get_model(role)` si role fourni
3. `OLLAMA_MODEL` global (fallback)