Spaces:
Running
Running
File size: 14,442 Bytes
01ae44e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 |
import os
import io
import json
import pathlib
import shutil
from typing import List, Tuple, Dict
import gradio as gr
import numpy as np
import faiss
from sentence_transformers import SentenceTransformer
from pypdf import PdfReader
import fitz # PyMuPDF
from collections import defaultdict
from openai import OpenAI
# =========================
# LLM Endpoint
# =========================
API_KEY = os.environ.get("API_KEY")
if not API_KEY:
raise RuntimeError("Missing API_KEY (set it in Hugging Face: Settings → Variables and secrets).")
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=API_KEY)
# Model configuration
# The model was hardcoded to "deepseek/deepseek-r1:free" as requested.
# The previous default was "Deepseek".
SINGLE_MODEL_NAME = "deepseek/deepseek-r1:free"
GEN_TEMPERATURE = 0.2
GEN_TOP_P = 0.95
GEN_MAX_TOKENS = 1024
EMB_MODEL_NAME = "intfloat/multilingual-e5-base"
def choose_store_dir() -> Tuple[str, bool]:
data_root = "/data"
if os.path.isdir(data_root) and os.access(data_root, os.W_OK):
d = os.path.join(data_root, "rag_store")
try:
os.makedirs(d, exist_ok=True)
testf = os.path.join(d, ".write_test")
with open(testf, "w", encoding="utf-8") as f:
f.write("ok")
os.remove(testf)
return d, True
except Exception:
pass
d = os.path.join(os.getcwd(), "store")
os.makedirs(d, exist_ok=True)
return d, False
STORE_DIR, IS_PERSISTENT = choose_store_dir()
META_PATH = os.path.join(STORE_DIR, "meta.json")
INDEX_PATH = os.path.join(STORE_DIR, "faiss.index")
LEGACY_STORE_DIR = os.path.join(os.getcwd(), "store")
def migrate_legacy_if_any():
try:
if IS_PERSISTENT:
legacy_meta = os.path.join(LEGACY_STORE_DIR, "meta.json")
legacy_index = os.path.join(LEGACY_STORE_DIR, "faiss.index")
if (not os.path.exists(META_PATH) or not os.path.exists(INDEX_PATH)) \
and os.path.isdir(LEGACY_STORE_DIR) \
and os.path.exists(legacy_meta) and os.path.exists(legacy_index):
shutil.copyfile(legacy_meta, META_PATH)
shutil.copyfile(legacy_index, INDEX_PATH)
except Exception:
pass
migrate_legacy_if_any()
_emb_model = None
_index: faiss.Index = None
_meta: Dict[str, Dict] = {}
DEFAULT_TOP_K = 6
DEFAULT_POOL_K = 40
DEFAULT_PER_SOURCE_CAP = 2
DEFAULT_STRATEGY = "mmr"
DEFAULT_MMR_LAMBDA = 0.5
def get_emb_model():
global _emb_model
if _emb_model is None:
_emb_model = SentenceTransformer(EMB_MODEL_NAME)
return _emb_model
def _ensure_index(dim: int):
global _index
if _index is None:
_index = faiss.IndexFlatIP(dim)
def _persist():
faiss.write_index(_index, INDEX_PATH)
with open(META_PATH, "w", encoding="utf-8") as f:
json.dump(_meta, f, ensure_ascii=False)
def _load_if_any():
global _index, _meta
if os.path.exists(INDEX_PATH) and os.path.exists(META_PATH):
_index = faiss.read_index(INDEX_PATH)
with open(META_PATH, "r", encoding="utf-8") as f:
_meta = json.load(f)
def _chunk_text(text: str, chunk_size: int = 800, overlap: int = 120) -> List[str]:
text = text.replace("\u0000", "")
res, i, n = [], 0, len(text)
while i < n:
j = min(i + chunk_size, n)
seg = text[i:j].strip()
if seg:
res.append(seg)
i = max(0, j - overlap)
if j >= n:
break
return res
def _read_bytes(file) -> bytes:
if isinstance(file, dict):
p = file.get("path") or file.get("name")
if p and os.path.exists(p):
with open(p, "rb") as f:
return f.read()
if "data" in file and isinstance(file["data"], (bytes, bytearray)):
return bytes(file["data"])
if isinstance(file, (str, pathlib.Path)):
with open(file, "rb") as f:
return f.read()
if hasattr(file, "read"):
try:
if hasattr(file, "seek"):
try:
file.seek(0)
except Exception:
pass
return file.read()
finally:
try:
file.close()
except Exception:
pass
raise ValueError("Unsupported file type from gr.File")
def _decode_best_effort(raw: bytes) -> str:
for enc in ["utf-8", "cp932", "shift_jis", "cp950", "big5", "gb18030", "latin-1"]:
try:
return raw.decode(enc)
except Exception:
continue
return raw.decode("utf-8", errors="ignore")
def _read_pdf(file_bytes: bytes) -> str:
try:
with fitz.open(stream=file_bytes, filetype="pdf") as doc:
if doc.is_encrypted:
try:
doc.authenticate("")
except Exception:
pass
texts = [(page.get_text("text") or "") for page in doc]
txt = "\n".join(texts)
if txt.strip():
return txt
except Exception:
pass
try:
reader = PdfReader(io.BytesIO(file_bytes))
pages = []
for p in reader.pages:
try:
pages.append(p.extract_text() or "")
except Exception:
pages.append("")
return "\n".join(pages)
except Exception:
return ""
def _read_any(file) -> str:
if isinstance(file, dict):
name = (file.get("orig_name") or file.get("name") or file.get("path") or "upload").lower()
else:
name = getattr(file, "name", None) or (str(file) if isinstance(file, (str, pathlib.Path)) else "upload")
name = name.lower()
raw = _read_bytes(file)
if name.endswith(".pdf"):
return _read_pdf(raw).replace("\u0000", "")
return _decode_best_effort(raw).replace("\u0000", "")
DOCS_DIR = os.path.join(os.getcwd(), "docs")
def get_docs_files() -> List[str]:
if not os.path.isdir(DOCS_DIR):
return []
files = []
for fname in os.listdir(DOCS_DIR):
if fname.lower().endswith((".pdf", ".txt")):
files.append(os.path.join(DOCS_DIR, fname))
return files
def build_corpus_from_docs():
global _index, _meta
files = get_docs_files()
if not files:
return "No files found in docs folder."
emb_model = get_emb_model()
chunks, sources, failed = [], [], []
_index = None
_meta = {}
for f in files:
fname = os.path.basename(f)
try:
text = _read_any(f) or ""
parts = _chunk_text(text)
if not parts:
failed.append(fname)
continue
chunks.extend(parts)
sources.extend([fname] * len(parts))
except Exception:
failed.append(fname)
if not chunks:
return "No text extracted from docs."
passages = [f"passage: {c}" for c in chunks]
vec = emb_model.encode(passages, batch_size=64, convert_to_numpy=True, normalize_embeddings=True)
_ensure_index(vec.shape[1])
_index.add(vec)
for i, (src, c) in enumerate(zip(sources, chunks)):
_meta[str(i)] = {"source": src, "text": c}
_persist()
msg = f"Indexed {len(chunks)} chunks from {len(files)} files."
if failed:
msg += f" Failed files: {', '.join(failed)}"
return msg
def _encode_query_vec(query: str) -> np.ndarray:
return get_emb_model().encode([f"query: {query}"], convert_to_numpy=True, normalize_embeddings=True)
def retrieve_candidates(qvec: np.ndarray, pool_k: int = 40) -> List[Tuple[str, float]]:
if _index is None or _index.ntotal == 0:
return []
pool_k = min(pool_k, _index.ntotal)
D, I = _index.search(qvec, pool_k)
return [(str(idx), float(score)) for idx, score in zip(I[0], D[0]) if idx != -1]
def select_diverse_by_source(cands: List[Tuple[str, float]], top_k: int = 6, per_source_cap: int = 2) -> List[Tuple[str, float]]:
if not cands:
return []
by_src: Dict[str, List[Tuple[str, float]]] = defaultdict(list)
for cid, s in cands:
m = _meta.get(cid)
if not m:
continue
by_src[m["source"]].append((cid, s))
for src in by_src:
by_src[src] = by_src[src][:per_source_cap]
picked, src_items, ptrs = [], [(s, it) for s, it in by_src.items()], {s: 0 for s in by_src}
while len(picked) < top_k:
advanced = False
for src, items in src_items:
i = ptrs[src]
if i < len(items):
picked.append(items[i])
ptrs[src] = i + 1
advanced = True
if len(picked) >= top_k:
break
if not advanced:
break
if len(picked) < top_k:
seen = {cid for cid, _ in picked}
for cid, s in cands:
if cid not in seen:
picked.append((cid, s))
seen.add(cid)
if len(picked) >= top_k:
break
return picked[:top_k]
def _encode_chunks_text(cids: List[str]) -> np.ndarray:
texts = [f"passage: {(_meta.get(cid) or {}).get('text','')}" for cid in cids]
return get_emb_model().encode(texts, convert_to_numpy=True, normalize_embeddings=True)
def select_diverse_mmr(cands: List[Tuple[str, float]], qvec: np.ndarray, top_k: int = 6, mmr_lambda: float = 0.5) -> List[Tuple[str, float]]:
if not cands:
return []
cids = [cid for cid, _ in cands]
cvecs = _encode_chunks_text(cids)
sim_to_q = (cvecs @ qvec.T).reshape(-1)
selected, remaining = [], set(range(len(cids)))
while len(selected) < min(top_k, len(cids)):
if not selected:
i = int(np.argmax(sim_to_q))
selected.append(i)
remaining.remove(i)
continue
S = cvecs[selected]
sim_to_S = (cvecs[list(remaining)] @ S.T)
max_sim_to_S = sim_to_S.max(axis=1) if sim_to_S.size > 0 else np.zeros((len(remaining),), dtype=np.float32)
sim_q_rem = sim_to_q[list(remaining)]
mmr_scores = mmr_lambda * sim_q_rem - (1.0 - mmr_lambda) * max_sim_to_S
j_rel = int(np.argmax(mmr_scores))
j = list(remaining)[j_rel]
selected.append(j)
remaining.remove(j)
return [(cids[i], float(sim_to_q[i])) for i in selected][:top_k]
def retrieve_diverse(query: str,
top_k: int = 6,
pool_k: int = 40,
per_source_cap: int = 2,
strategy: str = "mmr",
mmr_lambda: float = 0.5) -> List[Tuple[str, float]]:
qvec = _encode_query_vec(query)
cands = retrieve_candidates(qvec, pool_k=pool_k)
if strategy == "mmr":
return select_diverse_mmr(cands, qvec, top_k=top_k, mmr_lambda=mmr_lambda)
return select_diverse_by_source(cands, top_k=top_k, per_source_cap=per_source_cap)
def _format_ctx(hits: List[Tuple[str, float]]) -> str:
if not hits:
return ""
lines = []
for cid, _ in hits:
m = _meta.get(cid)
if not m:
continue
source_clean = m.get("source", "")
text_clean = (m.get("text", "") or "").replace("\n", " ")
lines.append(f"[{cid}] ({source_clean}) " + text_clean)
return "\n".join(lines[:10])
def chat_fn(message, history):
model_name = SINGLE_MODEL_NAME
if _index is None or _index.ntotal == 0:
status = build_corpus_from_docs()
if not (_index and _index.ntotal > 0):
yield f"**Index Status:** {status}\n\nPlease ensure you have a 'docs' folder with PDF/TXT files and try again."
return
hits = retrieve_diverse(
message,
top_k=6,
pool_k=40,
per_source_cap=2,
strategy="mmr",
mmr_lambda=0.5,
)
ctx = _format_ctx(hits) if hits else "(Current index is empty or no matching chunks found)"
sys_blocks = ["You are a research assistant who has an excellent factual understanding of the legal policies, regulations, and compliance of enterprises, governments, and global organizations. You are a research assistant who reads Legal papers and provides factual answers to queries. If you do not know the answer, you should convey that to the user instead of hallucinating. Answers must be based on retrieved content with evidence and source numbers cited. If retrieval is insufficient, please clearly explain the shortcomings. When answering, please cite the numbers, e.g., [3]"]
messages = [{"role": "system", "content": "\n\n".join(sys_blocks)}]
for u, a in history:
messages.append({"role": "user", "content": u})
messages.append({"role": "assistant", "content": a})
messages.append({"role": "user", "content": message})
try:
response = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=GEN_TEMPERATURE,
top_p=GEN_TOP_P,
max_tokens=GEN_MAX_TOKENS,
stream=True,
)
partial_message = ""
for chunk in response:
if hasattr(chunk.choices[0], "delta") and chunk.choices[0].delta.content is not None:
partial_message += chunk.choices[0].delta.content
yield partial_message
elif hasattr(chunk.choices[0], "message") and chunk.choices[0].message.content is not None:
partial_message += chunk.choices[0].message.content
yield partial_message
except Exception as e:
yield f"[Exception] {repr(e)}"
with gr.Blocks(theme=gr.themes.Default(primary_hue="sky")) as legalprodigy:
gr.Markdown("")
with gr.Row():
query_box = gr.Textbox(
placeholder="Try: Explain Arbiration Process",
scale=5
)
send_btn = gr.Button("Send", scale=1)
with gr.Row():
chatbot = gr.Chatbot(label="LegalProdigy")
state = gr.State([])
def chat_wrapper(user_message, history):
history = history or []
gen = chat_fn(user_message, history)
result = ""
for chunk in gen:
result = chunk
history.append((user_message, result))
return history, history
send_btn.click(
chat_wrapper,
inputs=[query_box, state],
outputs=[chatbot, state]
)
try:
_load_if_any()
except Exception:
pass
if __name__ == "__main__":
legalprodigy.launch()
|