Spaces:
Sleeping
Sleeping
Update app/api.py
Browse files- app/api.py +60 -89
app/api.py
CHANGED
@@ -1,12 +1,12 @@
|
|
1 |
# app/api.py
|
2 |
from __future__ import annotations
|
3 |
|
4 |
-
|
|
|
5 |
from collections import deque
|
6 |
-
from datetime import datetime
|
7 |
from time import perf_counter
|
8 |
-
import
|
9 |
-
import os
|
10 |
|
11 |
import faiss
|
12 |
from fastapi import FastAPI, UploadFile, File, HTTPException
|
@@ -16,11 +16,11 @@ from pydantic import BaseModel, Field
|
|
16 |
|
17 |
from .rag_system import SimpleRAG, UPLOAD_DIR, INDEX_DIR
|
18 |
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
app = FastAPI(title="RAG API", version="1.3.0")
|
23 |
|
|
|
24 |
app.add_middleware(
|
25 |
CORSMiddleware,
|
26 |
allow_origins=["*"],
|
@@ -31,9 +31,7 @@ app.add_middleware(
|
|
31 |
|
32 |
rag = SimpleRAG()
|
33 |
|
34 |
-
#
|
35 |
-
# Models
|
36 |
-
# ------------------------------------------------------------------------------
|
37 |
class UploadResponse(BaseModel):
|
38 |
filename: str
|
39 |
chunks_added: int
|
@@ -54,34 +52,28 @@ class HistoryResponse(BaseModel):
|
|
54 |
total_chunks: int
|
55 |
history: List[HistoryItem] = []
|
56 |
|
57 |
-
#
|
58 |
-
# Lightweight stats store (in-memory)
|
59 |
-
# ------------------------------------------------------------------------------
|
60 |
class StatsStore:
|
61 |
def __init__(self):
|
62 |
self.documents_indexed = 0
|
63 |
self.questions_answered = 0
|
64 |
self.latencies_ms = deque(maxlen=500)
|
65 |
-
|
66 |
-
self.
|
67 |
-
self.history = deque(maxlen=50) # recent questions
|
68 |
|
69 |
def add_docs(self, n: int):
|
70 |
if n > 0:
|
71 |
-
self.documents_indexed += n
|
72 |
|
73 |
def add_question(self, latency_ms: Optional[int] = None, q: Optional[str] = None):
|
74 |
self.questions_answered += 1
|
75 |
if latency_ms is not None:
|
76 |
self.latencies_ms.append(int(latency_ms))
|
77 |
-
if len(self.last7_questions)
|
78 |
-
self.last7_questions.appendleft(1)
|
79 |
-
else:
|
80 |
-
# attribute to "today" bucket
|
81 |
self.last7_questions[0] += 1
|
82 |
if q:
|
83 |
self.history.appendleft(
|
84 |
-
{"question": q, "timestamp": datetime.
|
85 |
)
|
86 |
|
87 |
@property
|
@@ -90,96 +82,76 @@ class StatsStore:
|
|
90 |
|
91 |
stats = StatsStore()
|
92 |
|
93 |
-
#
|
94 |
-
# Helpers
|
95 |
-
# ------------------------------------------------------------------------------
|
96 |
-
_GENERIC_PATTERNS = [
|
97 |
-
r"\bbased on document context\b",
|
98 |
-
r"\bappears to be\b",
|
99 |
-
r"\bgeneral (?:summary|overview)\b",
|
100 |
-
]
|
101 |
-
|
102 |
_STOPWORDS = {
|
103 |
"the","a","an","of","for","and","or","in","on","to","from","with","by","is","are",
|
104 |
-
"was","were","be","been","being","at","as","that","this","these","those","it",
|
105 |
-
"
|
106 |
}
|
107 |
|
108 |
-
def
|
|
|
|
|
|
|
109 |
if not text:
|
110 |
return True
|
111 |
low = text.strip().lower()
|
112 |
if len(low) < 15:
|
113 |
return True
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
return False
|
118 |
|
119 |
-
def
|
120 |
-
|
121 |
-
|
122 |
-
def extractive_answer(question: str, contexts: List[str], max_chars: int = 500) -> str:
|
123 |
-
"""
|
124 |
-
Simple keyword-based extractive fallback:
|
125 |
-
pick sentences containing most question tokens.
|
126 |
-
"""
|
127 |
if not contexts:
|
128 |
return "I couldn't find relevant information in the indexed documents for this question."
|
|
|
|
|
|
|
129 |
|
130 |
-
|
131 |
-
if not q_tokens:
|
132 |
-
# if question is e.g. numbers only
|
133 |
-
q_tokens = set(tokenize(" ".join(contexts[:1])))
|
134 |
-
|
135 |
-
# split into sentences
|
136 |
sentences: List[str] = []
|
137 |
for c in contexts:
|
138 |
-
|
139 |
-
# rough sentence split
|
140 |
-
for s in re.split(r"(?<=[\.!\?])\s+|\n+", c.strip()):
|
141 |
s = s.strip()
|
142 |
if s:
|
143 |
sentences.append(s)
|
144 |
|
145 |
-
if not sentences:
|
146 |
-
# fallback to first context chunk
|
147 |
-
return (contexts[0] or "")[:max_chars]
|
148 |
-
|
149 |
-
# score sentences
|
150 |
scored: List[tuple[int, str]] = []
|
151 |
for s in sentences:
|
152 |
-
|
153 |
-
|
154 |
-
scored.append((score, s))
|
155 |
-
|
156 |
-
# pick top sentences with score > 0, otherwise first few sentences
|
157 |
scored.sort(key=lambda x: (x[0], len(x[1])), reverse=True)
|
158 |
-
picked: List[str] = []
|
159 |
|
160 |
-
|
161 |
-
|
|
|
162 |
break
|
163 |
-
if len(" ".join(picked) + " " +
|
164 |
break
|
165 |
-
picked.append(
|
166 |
|
167 |
if not picked:
|
168 |
-
# no overlap, take first ~max_chars from contexts
|
169 |
return (contexts[0] or "")[:max_chars]
|
|
|
|
|
170 |
|
171 |
-
|
172 |
-
|
173 |
-
# ------------------------------------------------------------------------------
|
174 |
-
# Routes
|
175 |
-
# ------------------------------------------------------------------------------
|
176 |
@app.get("/")
|
177 |
def root():
|
178 |
return RedirectResponse(url="/docs")
|
179 |
|
180 |
@app.get("/health")
|
181 |
def health():
|
182 |
-
return {
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
@app.get("/debug/translate")
|
185 |
def debug_translate():
|
@@ -220,34 +192,35 @@ def ask_question(payload: AskRequest):
|
|
220 |
k = max(1, int(payload.top_k))
|
221 |
t0 = perf_counter()
|
222 |
|
223 |
-
#
|
224 |
try:
|
225 |
-
hits = rag.search(q, k=k) #
|
226 |
except Exception as e:
|
227 |
raise HTTPException(status_code=500, detail=f"Search failed: {e}")
|
228 |
|
229 |
-
contexts = [c for c, _ in (hits or []) if c] or (rag
|
230 |
|
231 |
if not contexts:
|
232 |
-
|
|
|
233 |
return AskResponse(
|
234 |
answer="I couldn't find relevant information in the indexed documents for this question.",
|
235 |
contexts=[]
|
236 |
)
|
237 |
|
238 |
-
#
|
239 |
try:
|
240 |
-
synthesized = rag.synthesize_answer(q, contexts) or ""
|
241 |
except Exception:
|
242 |
synthesized = ""
|
243 |
|
244 |
-
#
|
245 |
-
if
|
246 |
-
synthesized =
|
247 |
|
248 |
latency_ms = int((perf_counter() - t0) * 1000)
|
249 |
stats.add_question(latency_ms, q=q)
|
250 |
-
return AskResponse(answer=synthesized
|
251 |
|
252 |
@app.get("/get_history", response_model=HistoryResponse)
|
253 |
def get_history():
|
@@ -258,7 +231,6 @@ def get_history():
|
|
258 |
|
259 |
@app.get("/stats")
|
260 |
def stats_endpoint():
|
261 |
-
# keep backward compat fields + add dashboard-friendly metrics
|
262 |
return {
|
263 |
"documents_indexed": stats.documents_indexed,
|
264 |
"questions_answered": stats.questions_answered,
|
@@ -282,7 +254,6 @@ def reset_index():
|
|
282 |
os.remove(p)
|
283 |
except FileNotFoundError:
|
284 |
pass
|
285 |
-
# also reset stats counters to avoid stale analytics
|
286 |
stats.documents_indexed = 0
|
287 |
stats.questions_answered = 0
|
288 |
stats.latencies_ms.clear()
|
|
|
1 |
# app/api.py
|
2 |
from __future__ import annotations
|
3 |
|
4 |
+
import os
|
5 |
+
import re
|
6 |
from collections import deque
|
7 |
+
from datetime import datetime, timezone
|
8 |
from time import perf_counter
|
9 |
+
from typing import List, Optional, Dict, Any
|
|
|
10 |
|
11 |
import faiss
|
12 |
from fastapi import FastAPI, UploadFile, File, HTTPException
|
|
|
16 |
|
17 |
from .rag_system import SimpleRAG, UPLOAD_DIR, INDEX_DIR
|
18 |
|
19 |
+
__version__ = "1.3.1"
|
20 |
+
|
21 |
+
app = FastAPI(title="RAG API", version=__version__)
|
|
|
22 |
|
23 |
+
# CORS (Streamlit UI üçün)
|
24 |
app.add_middleware(
|
25 |
CORSMiddleware,
|
26 |
allow_origins=["*"],
|
|
|
31 |
|
32 |
rag = SimpleRAG()
|
33 |
|
34 |
+
# -------------------- Schemas --------------------
|
|
|
|
|
35 |
class UploadResponse(BaseModel):
|
36 |
filename: str
|
37 |
chunks_added: int
|
|
|
52 |
total_chunks: int
|
53 |
history: List[HistoryItem] = []
|
54 |
|
55 |
+
# -------------------- Stats (in-memory) --------------------
|
|
|
|
|
56 |
class StatsStore:
|
57 |
def __init__(self):
|
58 |
self.documents_indexed = 0
|
59 |
self.questions_answered = 0
|
60 |
self.latencies_ms = deque(maxlen=500)
|
61 |
+
self.last7_questions = deque([0] * 7, maxlen=7) # sadə günlük sayğac
|
62 |
+
self.history = deque(maxlen=50)
|
|
|
63 |
|
64 |
def add_docs(self, n: int):
|
65 |
if n > 0:
|
66 |
+
self.documents_indexed += int(n)
|
67 |
|
68 |
def add_question(self, latency_ms: Optional[int] = None, q: Optional[str] = None):
|
69 |
self.questions_answered += 1
|
70 |
if latency_ms is not None:
|
71 |
self.latencies_ms.append(int(latency_ms))
|
72 |
+
if len(self.last7_questions) == 7:
|
|
|
|
|
|
|
73 |
self.last7_questions[0] += 1
|
74 |
if q:
|
75 |
self.history.appendleft(
|
76 |
+
{"question": q, "timestamp": datetime.now(timezone.utc).isoformat(timespec="seconds")}
|
77 |
)
|
78 |
|
79 |
@property
|
|
|
82 |
|
83 |
stats = StatsStore()
|
84 |
|
85 |
+
# -------------------- Helpers --------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
_STOPWORDS = {
|
87 |
"the","a","an","of","for","and","or","in","on","to","from","with","by","is","are",
|
88 |
+
"was","were","be","been","being","at","as","that","this","these","those","it","its",
|
89 |
+
"into","than","then","so","such","about","over","per","via","vs","within"
|
90 |
}
|
91 |
|
92 |
+
def _tokenize(s: str) -> List[str]:
|
93 |
+
return [w for w in re.findall(r"[a-zA-Z0-9]+", s.lower()) if w and w not in _STOPWORDS and len(w) > 2]
|
94 |
+
|
95 |
+
def _is_generic_answer(text: str) -> bool:
|
96 |
if not text:
|
97 |
return True
|
98 |
low = text.strip().lower()
|
99 |
if len(low) < 15:
|
100 |
return True
|
101 |
+
# tipik generik pattern-lər
|
102 |
+
if "based on document context" in low or "appears to be" in low:
|
103 |
+
return True
|
104 |
return False
|
105 |
|
106 |
+
def _extractive_fallback(question: str, contexts: List[str], max_chars: int = 600) -> str:
|
107 |
+
""" Sualın açar sözlərinə əsasən kontekstdən cümlələr seç. """
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
if not contexts:
|
109 |
return "I couldn't find relevant information in the indexed documents for this question."
|
110 |
+
qtok = set(_tokenize(question))
|
111 |
+
if not qtok:
|
112 |
+
return (contexts[0] or "")[:max_chars]
|
113 |
|
114 |
+
# cümlələrə böl və skorla
|
|
|
|
|
|
|
|
|
|
|
115 |
sentences: List[str] = []
|
116 |
for c in contexts:
|
117 |
+
for s in re.split(r"(?<=[\.!\?])\s+|\n+", (c or "").strip()):
|
|
|
|
|
118 |
s = s.strip()
|
119 |
if s:
|
120 |
sentences.append(s)
|
121 |
|
|
|
|
|
|
|
|
|
|
|
122 |
scored: List[tuple[int, str]] = []
|
123 |
for s in sentences:
|
124 |
+
st = set(_tokenize(s))
|
125 |
+
scored.append((len(qtok & st), s))
|
|
|
|
|
|
|
126 |
scored.sort(key=lambda x: (x[0], len(x[1])), reverse=True)
|
|
|
127 |
|
128 |
+
picked: List[str] = []
|
129 |
+
for sc, s in scored:
|
130 |
+
if sc <= 0 and picked:
|
131 |
break
|
132 |
+
if len((" ".join(picked) + " " + s).strip()) > max_chars:
|
133 |
break
|
134 |
+
picked.append(s)
|
135 |
|
136 |
if not picked:
|
|
|
137 |
return (contexts[0] or "")[:max_chars]
|
138 |
+
bullets = "\n".join(f"- {p}" for p in picked)
|
139 |
+
return f"Answer (based on document context):\n{bullets}"
|
140 |
|
141 |
+
# -------------------- Routes --------------------
|
|
|
|
|
|
|
|
|
142 |
@app.get("/")
|
143 |
def root():
|
144 |
return RedirectResponse(url="/docs")
|
145 |
|
146 |
@app.get("/health")
|
147 |
def health():
|
148 |
+
return {
|
149 |
+
"status": "ok",
|
150 |
+
"version": app.version,
|
151 |
+
"summarizer": "extractive_en + translate + keyword_fallback",
|
152 |
+
"faiss_ntotal": int(getattr(rag.index, "ntotal", 0)),
|
153 |
+
"model_dim": int(getattr(rag.index, "d", rag.embed_dim)),
|
154 |
+
}
|
155 |
|
156 |
@app.get("/debug/translate")
|
157 |
def debug_translate():
|
|
|
192 |
k = max(1, int(payload.top_k))
|
193 |
t0 = perf_counter()
|
194 |
|
195 |
+
# 1) Həmişə sual embedding-i ilə axtar
|
196 |
try:
|
197 |
+
hits = rag.search(q, k=k) # List[Tuple[text, score]]
|
198 |
except Exception as e:
|
199 |
raise HTTPException(status_code=500, detail=f"Search failed: {e}")
|
200 |
|
201 |
+
contexts = [c for c, _ in (hits or []) if c] or (getattr(rag, "last_added", [])[:k] if getattr(rag, "last_added", None) else [])
|
202 |
|
203 |
if not contexts:
|
204 |
+
latency_ms = int((perf_counter() - t0) * 1000)
|
205 |
+
stats.add_question(latency_ms, q=q)
|
206 |
return AskResponse(
|
207 |
answer="I couldn't find relevant information in the indexed documents for this question.",
|
208 |
contexts=[]
|
209 |
)
|
210 |
|
211 |
+
# 2) Cavabı sintez et (rag içində LLM/rule-based ola bilər)
|
212 |
try:
|
213 |
+
synthesized = (rag.synthesize_answer(q, contexts) or "").strip()
|
214 |
except Exception:
|
215 |
synthesized = ""
|
216 |
|
217 |
+
# 3) Generic görünürsə, extractive fallback
|
218 |
+
if _is_generic_answer(synthesized):
|
219 |
+
synthesized = _extractive_fallback(q, contexts, max_chars=600)
|
220 |
|
221 |
latency_ms = int((perf_counter() - t0) * 1000)
|
222 |
stats.add_question(latency_ms, q=q)
|
223 |
+
return AskResponse(answer=synthesized, contexts=contexts)
|
224 |
|
225 |
@app.get("/get_history", response_model=HistoryResponse)
|
226 |
def get_history():
|
|
|
231 |
|
232 |
@app.get("/stats")
|
233 |
def stats_endpoint():
|
|
|
234 |
return {
|
235 |
"documents_indexed": stats.documents_indexed,
|
236 |
"questions_answered": stats.questions_answered,
|
|
|
254 |
os.remove(p)
|
255 |
except FileNotFoundError:
|
256 |
pass
|
|
|
257 |
stats.documents_indexed = 0
|
258 |
stats.questions_answered = 0
|
259 |
stats.latencies_ms.clear()
|