Spaces:
Sleeping
Sleeping
Speed up EXPLAIN: reduce LLM input and allow extractive-style explanation
Browse files- src/rag_core.py +197 -82
src/rag_core.py
CHANGED
|
@@ -2,13 +2,27 @@
|
|
| 2 |
# -*- coding: utf-8 -*-
|
| 3 |
|
| 4 |
"""
|
| 5 |
-
rag_core.py –
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
"""
|
| 7 |
|
| 8 |
import json
|
|
|
|
| 9 |
import re
|
| 10 |
from pathlib import Path
|
| 11 |
-
from typing import List,
|
| 12 |
|
| 13 |
from langchain_community.vectorstores import FAISS
|
| 14 |
from langchain_huggingface import HuggingFaceEmbeddings
|
|
@@ -22,43 +36,55 @@ DB_DIR = Path("db/faiss_code_edu_by_article")
|
|
| 22 |
|
| 23 |
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
|
| 24 |
|
| 25 |
-
TOP_K_FETCH = 30
|
| 26 |
TOP_K_FINAL = 3
|
| 27 |
-
SCORE_THRESHOLD = 1.10
|
| 28 |
|
| 29 |
-
MAX_CHARS_PER_DOC = 1200
|
| 30 |
SNIPPET_CHARS = 260
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
ARTICLE_ID_RE = re.compile(
|
| 33 |
r"\b(?:article\s+)?([LDR]\s?\d{1,4}(?:[.-]\d+){0,4})\b",
|
| 34 |
-
flags=re.IGNORECASE
|
| 35 |
)
|
| 36 |
|
| 37 |
EXPLAIN_TRIGGERS = [
|
| 38 |
-
"explique", "
|
| 39 |
-
"
|
|
|
|
| 40 |
]
|
| 41 |
|
| 42 |
LIST_TRIGGERS = [
|
| 43 |
-
"quels articles", "
|
|
|
|
|
|
|
| 44 |
]
|
| 45 |
|
| 46 |
FULLTEXT_TRIGGERS = [
|
| 47 |
-
"
|
|
|
|
|
|
|
|
|
|
| 48 |
]
|
| 49 |
|
| 50 |
_REFUSAL = "Je ne peux pas répondre avec certitude à partir des articles fournis."
|
| 51 |
_EXPLAIN_REFUSAL = (
|
| 52 |
-
"Pour expliquer
|
| 53 |
-
"
|
| 54 |
)
|
| 55 |
|
|
|
|
| 56 |
|
| 57 |
-
#
|
| 58 |
-
|
| 59 |
llm = Llama(
|
| 60 |
model_path="models/model.gguf",
|
| 61 |
-
n_ctx=
|
| 62 |
n_threads=10,
|
| 63 |
n_batch=128,
|
| 64 |
verbose=False,
|
|
@@ -66,7 +92,6 @@ llm = Llama(
|
|
| 66 |
|
| 67 |
|
| 68 |
def llm_generate_qa(prompt: str) -> str:
|
| 69 |
-
"""Réponse courte, stricte"""
|
| 70 |
out = llm.create_chat_completion(
|
| 71 |
messages=[{"role": "user", "content": prompt}],
|
| 72 |
temperature=0.1,
|
|
@@ -75,12 +100,15 @@ def llm_generate_qa(prompt: str) -> str:
|
|
| 75 |
return out["choices"][0]["message"]["content"].strip()
|
| 76 |
|
| 77 |
|
| 78 |
-
def
|
| 79 |
-
"""
|
|
|
|
|
|
|
|
|
|
| 80 |
out = llm.create_chat_completion(
|
| 81 |
messages=[{"role": "user", "content": prompt}],
|
| 82 |
temperature=0.2,
|
| 83 |
-
max_tokens=
|
| 84 |
)
|
| 85 |
return out["choices"][0]["message"]["content"].strip()
|
| 86 |
|
|
@@ -92,31 +120,42 @@ def normalize_article_id(raw: str) -> str:
|
|
| 92 |
|
| 93 |
|
| 94 |
def extract_article_id(q: str) -> Optional[str]:
|
| 95 |
-
m = ARTICLE_ID_RE.search(q)
|
| 96 |
return normalize_article_id(m.group(1)) if m else None
|
| 97 |
|
| 98 |
|
| 99 |
def is_explain_request(q: str) -> bool:
|
| 100 |
-
ql = q.lower()
|
| 101 |
return any(t in ql for t in EXPLAIN_TRIGGERS)
|
| 102 |
|
| 103 |
|
| 104 |
def is_list_request(q: str) -> bool:
|
| 105 |
-
ql = q.lower()
|
| 106 |
return any(t in ql for t in LIST_TRIGGERS)
|
| 107 |
|
| 108 |
|
| 109 |
def is_fulltext_request(q: str) -> bool:
|
| 110 |
-
ql = q.lower()
|
| 111 |
return any(t in ql for t in FULLTEXT_TRIGGERS)
|
| 112 |
|
| 113 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
def load_article_text(article_id: str) -> Optional[str]:
|
|
|
|
|
|
|
|
|
|
| 115 |
with CHUNKS_PATH.open("r", encoding="utf-8") as f:
|
| 116 |
for line in f:
|
|
|
|
|
|
|
| 117 |
obj = json.loads(line)
|
| 118 |
-
|
| 119 |
-
|
|
|
|
| 120 |
return None
|
| 121 |
|
| 122 |
|
|
@@ -132,33 +171,120 @@ def get_vectorstore() -> FAISS:
|
|
| 132 |
_VS = FAISS.load_local(
|
| 133 |
str(DB_DIR),
|
| 134 |
embeddings,
|
| 135 |
-
allow_dangerous_deserialization=True
|
| 136 |
)
|
| 137 |
return _VS
|
| 138 |
|
| 139 |
|
| 140 |
-
# ====================
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
-
def
|
| 143 |
-
|
| 144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
|
| 146 |
-
ARTICLE :
|
| 147 |
-
[{article_id}]
|
| 148 |
-
{article_text}
|
| 149 |
|
| 150 |
-
|
| 151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
- Pas de généralisation
|
| 156 |
-
- Pas de suppositions
|
| 157 |
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
- Ton clair et accessible
|
| 161 |
-
- Aucune citation d'autres articles
|
| 162 |
"""
|
| 163 |
|
| 164 |
|
|
@@ -180,75 +306,64 @@ CONTEXTE :
|
|
| 180 |
FORMAT FINAL :
|
| 181 |
Réponse courte.
|
| 182 |
Dernière ligne : Articles cités : A, B
|
| 183 |
-
"""
|
| 184 |
|
| 185 |
|
| 186 |
# ==================== CORE ====================
|
| 187 |
|
| 188 |
def answer_query(q: str) -> Dict[str, Any]:
|
| 189 |
-
q = q.strip()
|
| 190 |
if not q:
|
| 191 |
return {"mode": "QA", "answer": _REFUSAL, "articles": []}
|
| 192 |
|
| 193 |
article_id = extract_article_id(q)
|
| 194 |
|
| 195 |
-
# ---------- EXPLAIN ----------
|
| 196 |
if is_explain_request(q):
|
| 197 |
if not article_id:
|
| 198 |
-
return {
|
| 199 |
-
"mode": "EXPLAIN",
|
| 200 |
-
"answer": _EXPLAIN_REFUSAL,
|
| 201 |
-
"articles": []
|
| 202 |
-
}
|
| 203 |
|
| 204 |
text = load_article_text(article_id)
|
| 205 |
if not text:
|
| 206 |
-
return {
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 220 |
|
| 221 |
# ---------- FULLTEXT ----------
|
| 222 |
if article_id and is_fulltext_request(q):
|
| 223 |
text = load_article_text(article_id)
|
| 224 |
-
return {
|
| 225 |
-
"mode": "FULLTEXT",
|
| 226 |
-
"answer": text or _REFUSAL,
|
| 227 |
-
"articles": [article_id]
|
| 228 |
-
}
|
| 229 |
|
| 230 |
# ---------- LIST ----------
|
| 231 |
if is_list_request(q):
|
| 232 |
vs = get_vectorstore()
|
| 233 |
docs = vs.similarity_search(q, k=5)
|
| 234 |
-
arts = list({normalize_article_id(d.metadata
|
| 235 |
-
return {
|
| 236 |
-
"mode": "LIST",
|
| 237 |
-
"answer": "",
|
| 238 |
-
"articles": arts
|
| 239 |
-
}
|
| 240 |
|
| 241 |
# ---------- QA ----------
|
| 242 |
vs = get_vectorstore()
|
| 243 |
docs = vs.similarity_search(q, k=TOP_K_FINAL)
|
| 244 |
context = "\n\n".join(d.page_content for d in docs)
|
| 245 |
-
articles = [normalize_article_id(d.metadata
|
| 246 |
|
| 247 |
prompt = build_qa_prompt(q, context, articles)
|
| 248 |
answer = llm_generate_qa(prompt)
|
| 249 |
|
| 250 |
-
return {
|
| 251 |
-
"mode": "QA",
|
| 252 |
-
"answer": answer,
|
| 253 |
-
"articles": articles
|
| 254 |
-
}
|
|
|
|
| 2 |
# -*- coding: utf-8 -*-
|
| 3 |
|
| 4 |
"""
|
| 5 |
+
rag_core.py – EXPLAIN ultra rapide via résumé extractif (text mining)
|
| 6 |
+
|
| 7 |
+
Objectif :
|
| 8 |
+
- LIST & FULLTEXT restent instantanés (pas de LLM)
|
| 9 |
+
- EXPLAIN devient très rapide : extraction de 3–6 segments clés de l’article
|
| 10 |
+
- QA reste possible (LLM), mais lent (CPU)
|
| 11 |
+
|
| 12 |
+
Principe EXPLAIN :
|
| 13 |
+
- ID d’article obligatoire, sinon refus.
|
| 14 |
+
- On charge le texte exact de l’article depuis chunks_articles.jsonl
|
| 15 |
+
- On produit une "explication" par extraction (aucune génération) -> zéro hallucination
|
| 16 |
+
- Optionnel : reformulation LLM sur le résumé (désactivé par défaut)
|
| 17 |
+
|
| 18 |
+
Ce fichier remplace le précédent (qui envoyait l’article intégral au LLM en EXPLAIN).
|
| 19 |
"""
|
| 20 |
|
| 21 |
import json
|
| 22 |
+
import os
|
| 23 |
import re
|
| 24 |
from pathlib import Path
|
| 25 |
+
from typing import List, Optional, Dict, Any
|
| 26 |
|
| 27 |
from langchain_community.vectorstores import FAISS
|
| 28 |
from langchain_huggingface import HuggingFaceEmbeddings
|
|
|
|
| 36 |
|
| 37 |
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
|
| 38 |
|
|
|
|
| 39 |
TOP_K_FINAL = 3
|
|
|
|
| 40 |
|
|
|
|
| 41 |
SNIPPET_CHARS = 260
|
| 42 |
|
| 43 |
+
# --- Résumé extractif ---
|
| 44 |
+
EXTRACT_MAX_SEGMENTS = 5 # nb max de segments extraits
|
| 45 |
+
EXTRACT_MAX_CHARS_TOTAL = 900 # garde-fou (résumé total)
|
| 46 |
+
EXTRACT_MIN_SEG_LEN = 30 # ignore segments trop courts
|
| 47 |
+
EXTRACT_MAX_SEG_LEN = 420 # tronque segments trop longs
|
| 48 |
+
|
| 49 |
+
# option : reformulation LLM sur résumé extractif (OFF par défaut)
|
| 50 |
+
EXPLAIN_USE_LLM = os.environ.get("EXPLAIN_USE_LLM", "0").strip() == "1"
|
| 51 |
+
|
| 52 |
ARTICLE_ID_RE = re.compile(
|
| 53 |
r"\b(?:article\s+)?([LDR]\s?\d{1,4}(?:[.-]\d+){0,4})\b",
|
| 54 |
+
flags=re.IGNORECASE,
|
| 55 |
)
|
| 56 |
|
| 57 |
EXPLAIN_TRIGGERS = [
|
| 58 |
+
"explique", "expliquer", "explication",
|
| 59 |
+
"résume", "resume", "résumé", "reformule", "simplifie",
|
| 60 |
+
"en termes simples", "vulgarise", "clarifie",
|
| 61 |
]
|
| 62 |
|
| 63 |
LIST_TRIGGERS = [
|
| 64 |
+
"quels articles", "quelles dispositions", "articles parlent",
|
| 65 |
+
"articles qui parlent", "articles sur", "donne les articles",
|
| 66 |
+
"cite les articles", "références", "references",
|
| 67 |
]
|
| 68 |
|
| 69 |
FULLTEXT_TRIGGERS = [
|
| 70 |
+
"contenu exact", "texte exact", "texte intégral", "texte integral",
|
| 71 |
+
"intégral", "integral", "cite intégralement", "cite integralement",
|
| 72 |
+
"donne l'intégralité", "donne l'integralite", "recopie", "reproduis",
|
| 73 |
+
"affiche l'article", "donne l'article", "donne moi l'article",
|
| 74 |
]
|
| 75 |
|
| 76 |
_REFUSAL = "Je ne peux pas répondre avec certitude à partir des articles fournis."
|
| 77 |
_EXPLAIN_REFUSAL = (
|
| 78 |
+
"Pour expliquer ou résumer, indique un identifiant d’article (ex : D422-5). "
|
| 79 |
+
"Sinon, commence par : \"Quels articles parlent de … ?\""
|
| 80 |
)
|
| 81 |
|
| 82 |
+
# ==================== LLM INIT (QA + option EXPLAIN LLM) ====================
|
| 83 |
|
| 84 |
+
# Le LLM est utile pour QA. Pour EXPLAIN "très vite", on le désactive par défaut.
|
|
|
|
| 85 |
llm = Llama(
|
| 86 |
model_path="models/model.gguf",
|
| 87 |
+
n_ctx=1024, # réduit pour CPU
|
| 88 |
n_threads=10,
|
| 89 |
n_batch=128,
|
| 90 |
verbose=False,
|
|
|
|
| 92 |
|
| 93 |
|
| 94 |
def llm_generate_qa(prompt: str) -> str:
|
|
|
|
| 95 |
out = llm.create_chat_completion(
|
| 96 |
messages=[{"role": "user", "content": prompt}],
|
| 97 |
temperature=0.1,
|
|
|
|
| 100 |
return out["choices"][0]["message"]["content"].strip()
|
| 101 |
|
| 102 |
|
| 103 |
+
def llm_generate_explain_from_summary(prompt: str) -> str:
|
| 104 |
+
"""
|
| 105 |
+
Reformulation optionnelle du résumé extractif.
|
| 106 |
+
On reste court pour ne pas exploser la latence CPU.
|
| 107 |
+
"""
|
| 108 |
out = llm.create_chat_completion(
|
| 109 |
messages=[{"role": "user", "content": prompt}],
|
| 110 |
temperature=0.2,
|
| 111 |
+
max_tokens=160,
|
| 112 |
)
|
| 113 |
return out["choices"][0]["message"]["content"].strip()
|
| 114 |
|
|
|
|
| 120 |
|
| 121 |
|
| 122 |
def extract_article_id(q: str) -> Optional[str]:
|
| 123 |
+
m = ARTICLE_ID_RE.search(q or "")
|
| 124 |
return normalize_article_id(m.group(1)) if m else None
|
| 125 |
|
| 126 |
|
| 127 |
def is_explain_request(q: str) -> bool:
|
| 128 |
+
ql = (q or "").lower()
|
| 129 |
return any(t in ql for t in EXPLAIN_TRIGGERS)
|
| 130 |
|
| 131 |
|
| 132 |
def is_list_request(q: str) -> bool:
|
| 133 |
+
ql = (q or "").lower()
|
| 134 |
return any(t in ql for t in LIST_TRIGGERS)
|
| 135 |
|
| 136 |
|
| 137 |
def is_fulltext_request(q: str) -> bool:
|
| 138 |
+
ql = (q or "").lower()
|
| 139 |
return any(t in ql for t in FULLTEXT_TRIGGERS)
|
| 140 |
|
| 141 |
|
| 142 |
+
def safe_snippet(text: str, n: int) -> str:
|
| 143 |
+
t = " ".join((text or "").split())
|
| 144 |
+
return t if len(t) <= n else t[:n].rstrip() + "…"
|
| 145 |
+
|
| 146 |
+
|
| 147 |
def load_article_text(article_id: str) -> Optional[str]:
|
| 148 |
+
if not CHUNKS_PATH.exists():
|
| 149 |
+
raise FileNotFoundError(f"Fichier chunks introuvable : {CHUNKS_PATH}")
|
| 150 |
+
|
| 151 |
with CHUNKS_PATH.open("r", encoding="utf-8") as f:
|
| 152 |
for line in f:
|
| 153 |
+
if not line.strip():
|
| 154 |
+
continue
|
| 155 |
obj = json.loads(line)
|
| 156 |
+
aid = normalize_article_id(obj.get("article_id", ""))
|
| 157 |
+
if aid == article_id:
|
| 158 |
+
return (obj.get("text") or "").strip()
|
| 159 |
return None
|
| 160 |
|
| 161 |
|
|
|
|
| 171 |
_VS = FAISS.load_local(
|
| 172 |
str(DB_DIR),
|
| 173 |
embeddings,
|
| 174 |
+
allow_dangerous_deserialization=True,
|
| 175 |
)
|
| 176 |
return _VS
|
| 177 |
|
| 178 |
|
| 179 |
+
# ==================== EXTRACTIVE SUMMARY (FAST) ====================
|
| 180 |
+
|
| 181 |
+
_NORMATIVE_PATTERNS = [
|
| 182 |
+
# Verbes normatifs / obligations
|
| 183 |
+
r"\bdoit\b", r"\bdoivent\b", r"\best\b", r"\bsont\b",
|
| 184 |
+
r"\bpeut\b", r"\bpeuvent\b",
|
| 185 |
+
r"\best tenu\b", r"\bsont tenus\b", r"\best tenu de\b",
|
| 186 |
+
r"\best interdit\b", r"\bsont interdits\b", r"\bil est interdit\b",
|
| 187 |
+
r"\bobligatoire\b", r"\bobligation\b",
|
| 188 |
+
# Conditions / exceptions
|
| 189 |
+
r"\bsi\b", r"\blorsque\b", r"\bsauf\b", r"\bà condition\b", r"\ba condition\b",
|
| 190 |
+
r"\bdans le cas\b", r"\ben cas\b", r"\btoutefois\b",
|
| 191 |
+
# Structure
|
| 192 |
+
r"\bI\.\b", r"\bII\.\b", r"\bIII\.\b", r"\b1°\b", r"\b2°\b", r"\b3°\b",
|
| 193 |
+
]
|
| 194 |
+
|
| 195 |
|
| 196 |
+
def _split_into_segments(text: str) -> List[str]:
|
| 197 |
+
"""
|
| 198 |
+
Découpe grossière mais robuste pour du juridique :
|
| 199 |
+
- on coupe par lignes / alinéas
|
| 200 |
+
- puis on recoupe si lignes trop longues via ; .
|
| 201 |
+
"""
|
| 202 |
+
if not text:
|
| 203 |
+
return []
|
| 204 |
+
|
| 205 |
+
# 1) alinéas
|
| 206 |
+
lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
|
| 207 |
+
segs: List[str] = []
|
| 208 |
+
for ln in lines:
|
| 209 |
+
# 2) recoupe douce
|
| 210 |
+
if len(ln) > 600:
|
| 211 |
+
parts = re.split(r"(?<=[.;:])\s+", ln)
|
| 212 |
+
segs.extend([p.strip() for p in parts if p.strip()])
|
| 213 |
+
else:
|
| 214 |
+
segs.append(ln)
|
| 215 |
+
return segs
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
def _score_segment(seg: str) -> int:
|
| 219 |
+
s = 0
|
| 220 |
+
low = seg.lower()
|
| 221 |
+
for pat in _NORMATIVE_PATTERNS:
|
| 222 |
+
if re.search(pat, low, flags=re.IGNORECASE):
|
| 223 |
+
s += 2
|
| 224 |
+
# bonus si segment contient des marqueurs juridiques
|
| 225 |
+
if re.search(r"\b(décret|arrêté|loi|code)\b", low):
|
| 226 |
+
s += 1
|
| 227 |
+
# pénalité si segment trop long (moins lisible)
|
| 228 |
+
if len(seg) > 450:
|
| 229 |
+
s -= 1
|
| 230 |
+
return s
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
def extractive_explain(article_id: str, article_text: str) -> str:
|
| 234 |
+
"""
|
| 235 |
+
Produit une 'explication' très rapide :
|
| 236 |
+
- sélection de segments clés (extraction)
|
| 237 |
+
- aucune génération => zéro hallucination
|
| 238 |
+
"""
|
| 239 |
+
segs = _split_into_segments(article_text)
|
| 240 |
+
cleaned = []
|
| 241 |
+
for s in segs:
|
| 242 |
+
s = " ".join(s.split())
|
| 243 |
+
if len(s) < EXTRACT_MIN_SEG_LEN:
|
| 244 |
+
continue
|
| 245 |
+
if len(s) > EXTRACT_MAX_SEG_LEN:
|
| 246 |
+
s = s[:EXTRACT_MAX_SEG_LEN].rstrip() + "…"
|
| 247 |
+
cleaned.append(s)
|
| 248 |
+
|
| 249 |
+
if not cleaned:
|
| 250 |
+
return f"Résumé impossible : texte vide ou non exploitable.\n\nArticles cités : {article_id}"
|
| 251 |
+
|
| 252 |
+
scored = sorted((( _score_segment(s), s) for s in cleaned), key=lambda x: x[0], reverse=True)
|
| 253 |
+
|
| 254 |
+
# garde ceux qui ont un score positif, sinon fallback sur les premiers segments
|
| 255 |
+
picked = [s for (sc, s) in scored if sc > 0][:EXTRACT_MAX_SEGMENTS]
|
| 256 |
+
if not picked:
|
| 257 |
+
picked = cleaned[:min(EXTRACT_MAX_SEGMENTS, len(cleaned))]
|
| 258 |
+
|
| 259 |
+
# garde-fou longueur totale
|
| 260 |
+
out_parts = []
|
| 261 |
+
total = 0
|
| 262 |
+
for s in picked:
|
| 263 |
+
if total + len(s) > EXTRACT_MAX_CHARS_TOTAL and out_parts:
|
| 264 |
+
break
|
| 265 |
+
out_parts.append(f"- {s}")
|
| 266 |
+
total += len(s)
|
| 267 |
+
|
| 268 |
+
body = (
|
| 269 |
+
"Points clés (extraction du texte, sans reformulation) :\n"
|
| 270 |
+
+ "\n".join(out_parts)
|
| 271 |
+
)
|
| 272 |
+
return f"{body}\n\nArticles cités : {article_id}"
|
| 273 |
|
|
|
|
|
|
|
|
|
|
| 274 |
|
| 275 |
+
def build_explain_llm_prompt(article_id: str, extractive_summary: str) -> str:
|
| 276 |
+
"""
|
| 277 |
+
Reformulation LLM optionnelle sur RÉSUMÉ COURT (pas sur l’article int��gral).
|
| 278 |
+
"""
|
| 279 |
+
return f"""Tu es un assistant pédagogique. Tu dois reformuler en termes simples le contenu fourni.
|
| 280 |
+
Interdictions : rien inventer, rien ajouter, pas d’autres articles.
|
| 281 |
+
Tu dois rester fidèle aux points ci-dessous.
|
| 282 |
|
| 283 |
+
CONTENU (extrait du texte) :
|
| 284 |
+
{extractive_summary}
|
|
|
|
|
|
|
| 285 |
|
| 286 |
+
Donne une explication en 4–6 phrases maximum.
|
| 287 |
+
Dernière ligne : Articles cités : {article_id}
|
|
|
|
|
|
|
| 288 |
"""
|
| 289 |
|
| 290 |
|
|
|
|
| 306 |
FORMAT FINAL :
|
| 307 |
Réponse courte.
|
| 308 |
Dernière ligne : Articles cités : A, B
|
| 309 |
+
""".strip()
|
| 310 |
|
| 311 |
|
| 312 |
# ==================== CORE ====================
|
| 313 |
|
| 314 |
def answer_query(q: str) -> Dict[str, Any]:
|
| 315 |
+
q = (q or "").strip()
|
| 316 |
if not q:
|
| 317 |
return {"mode": "QA", "answer": _REFUSAL, "articles": []}
|
| 318 |
|
| 319 |
article_id = extract_article_id(q)
|
| 320 |
|
| 321 |
+
# ---------- EXPLAIN (FAST) ----------
|
| 322 |
if is_explain_request(q):
|
| 323 |
if not article_id:
|
| 324 |
+
return {"mode": "EXPLAIN", "answer": _EXPLAIN_REFUSAL, "articles": []}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 325 |
|
| 326 |
text = load_article_text(article_id)
|
| 327 |
if not text:
|
| 328 |
+
return {"mode": "EXPLAIN", "answer": f"Article {article_id} introuvable.", "articles": []}
|
| 329 |
+
|
| 330 |
+
# 1) explication immédiate par extraction (très rapide)
|
| 331 |
+
extractive = extractive_explain(article_id, text)
|
| 332 |
+
|
| 333 |
+
# 2) optionnel : mini reformulation LLM sur le résumé (pas sur l’article)
|
| 334 |
+
if EXPLAIN_USE_LLM:
|
| 335 |
+
try:
|
| 336 |
+
prompt = build_explain_llm_prompt(article_id, extractive)
|
| 337 |
+
llm_ans = llm_generate_explain_from_summary(prompt).strip()
|
| 338 |
+
# garantie citation
|
| 339 |
+
if "Articles cités" not in llm_ans:
|
| 340 |
+
llm_ans = llm_ans.rstrip() + f"\n\nArticles cités : {article_id}"
|
| 341 |
+
return {"mode": "EXPLAIN", "answer": llm_ans, "articles": [article_id]}
|
| 342 |
+
except Exception:
|
| 343 |
+
# fallback extractif si souci LLM
|
| 344 |
+
return {"mode": "EXPLAIN", "answer": extractive, "articles": [article_id]}
|
| 345 |
+
|
| 346 |
+
return {"mode": "EXPLAIN", "answer": extractive, "articles": [article_id]}
|
| 347 |
|
| 348 |
# ---------- FULLTEXT ----------
|
| 349 |
if article_id and is_fulltext_request(q):
|
| 350 |
text = load_article_text(article_id)
|
| 351 |
+
return {"mode": "FULLTEXT", "answer": text or _REFUSAL, "articles": [article_id]}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 352 |
|
| 353 |
# ---------- LIST ----------
|
| 354 |
if is_list_request(q):
|
| 355 |
vs = get_vectorstore()
|
| 356 |
docs = vs.similarity_search(q, k=5)
|
| 357 |
+
arts = list({normalize_article_id(d.metadata.get("article_id", "")) for d in docs})
|
| 358 |
+
return {"mode": "LIST", "answer": "", "articles": arts}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
|
| 360 |
# ---------- QA ----------
|
| 361 |
vs = get_vectorstore()
|
| 362 |
docs = vs.similarity_search(q, k=TOP_K_FINAL)
|
| 363 |
context = "\n\n".join(d.page_content for d in docs)
|
| 364 |
+
articles = [normalize_article_id(d.metadata.get("article_id", "")) for d in docs]
|
| 365 |
|
| 366 |
prompt = build_qa_prompt(q, context, articles)
|
| 367 |
answer = llm_generate_qa(prompt)
|
| 368 |
|
| 369 |
+
return {"mode": "QA", "answer": answer, "articles": articles}
|
|
|
|
|
|
|
|
|
|
|
|