CAIAssignmentGradio / generate_indexes.py
rishabhsetiya's picture
Create generate_indexes.py
c69001b verified
raw
history blame
5.64 kB
import os
import re
import json
import pickle
from typing import List, Dict
import numpy as np
import faiss
import pandas as pd
import tabula
from sentence_transformers import SentenceTransformer
from rank_bm25 import BM25Okapi
# ---------------- Config ----------------
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
PDF_PATH = "MakeMyTrip_Financial_Statements.pdf"
OUT_DIR = "data/index_merged"
# Paths for saved chunks & indices
CHUNKS_100_PATH = os.path.join(OUT_DIR, "chunks_100.json")
CHUNKS_400_PATH = os.path.join(OUT_DIR, "chunks_400.json")
CHUNKS_MERGED_PATH = os.path.join(OUT_DIR, "chunks_merged.json")
FAISS_PATH = os.path.join(OUT_DIR, "faiss_merged.index")
BM25_PATH = os.path.join(OUT_DIR, "bm25_merged.pkl")
META_PATH = os.path.join(OUT_DIR, "meta_merged.pkl")
# ---------------- Utils ----------------
_tok_pat = re.compile(r"[a-z0-9]+", re.I)
def simple_tokenize(text: str):
return _tok_pat.findall((text or "").lower())
def create_chunks(texts: List[str], max_tokens: int) -> List[str]:
"""Simple word-based tokenizer to split texts into chunks."""
chunks, current_chunk, current_tokens = [], [], 0
for text in texts:
tokens = re.findall(r"\w+", text)
if current_tokens + len(tokens) > max_tokens:
chunks.append(" ".join(current_chunk))
current_chunk, current_tokens = [], 0
current_chunk.append(text)
current_tokens += len(tokens)
if current_chunk:
chunks.append(" ".join(current_chunk))
return chunks
def extract_tables_from_pdf(pdf_path: str, pages="all") -> List[Dict]:
"""Extract tables from financial PDF into structured row-year-value dicts."""
tables = tabula.read_pdf(
pdf_path,
pages=pages,
multiple_tables=True,
pandas_options={'dtype': str}
)
table_rows = []
row_id = 0
for df in tables:
if df.empty:
continue
df = df.replace(r'\n', ' ', regex=True).fillna("")
headers = list(df.iloc[0])
if any(re.match(r"20\d{2}", str(c)) for c in headers):
df.columns = [c.strip() for c in headers]
df = df.drop(0).reset_index(drop=True)
for _, row in df.iterrows():
metric = str(row.iloc[0]).strip()
if not metric or metric.lower() in ["note", ""]:
continue
values = {}
for col, val in row.items():
if re.match(r"20\d{2}", str(col)):
clean_val = str(val).replace(",", "").strip()
if clean_val and clean_val not in ["-", "—", "nan"]:
values[str(col)] = clean_val
if not values:
continue
table_rows.append({
"id": f"table-{row_id}",
"metric": metric,
"years": list(values.keys()),
"values": values,
"content": f"{metric} values: {json.dumps(values)}",
"source": "table"
})
row_id += 1
print(f"Extracted {len(table_rows)} rows from PDF tables")
return table_rows
def build_dense_faiss(texts: List[str], out_path: str):
print(f"Embedding {len(texts)} docs with {EMBED_MODEL} ...")
model = SentenceTransformer(EMBED_MODEL)
emb = model.encode(texts, convert_to_numpy=True, batch_size=64, show_progress_bar=True)
faiss.normalize_L2(emb)
dim = emb.shape[1]
index = faiss.IndexFlatIP(dim)
index.add(emb)
faiss.write_index(index, out_path)
print(f"FAISS index built & saved -> {out_path}")
def build_bm25(texts: List[str], out_path: str):
tokenized = [simple_tokenize(t) for t in texts]
bm25 = BM25Okapi(tokenized)
with open(out_path, "wb") as f:
pickle.dump({"bm25": bm25, "tokenized_corpus": tokenized}, f)
print(f"BM25 index built & saved -> {out_path}")
# ---------------- Main ----------------
def main():
os.makedirs(OUT_DIR, exist_ok=True)
# 1) Extract table rows
docs = extract_tables_from_pdf(PDF_PATH, pages="all")
all_texts = [d["content"] for d in docs]
# 2) Create chunks of size 100 and 400
chunks_100 = create_chunks(all_texts, 100)
chunks_400 = create_chunks(all_texts, 400)
# 3) Save them separately
with open(CHUNKS_100_PATH, "w", encoding="utf-8") as f:
json.dump(chunks_100, f, indent=2, ensure_ascii=False)
with open(CHUNKS_400_PATH, "w", encoding="utf-8") as f:
json.dump(chunks_400, f, indent=2, ensure_ascii=False)
print(f"Saved {len(chunks_100)} chunks_100 -> {CHUNKS_100_PATH}")
print(f"Saved {len(chunks_400)} chunks_400 -> {CHUNKS_400_PATH}")
# 4) Merge with metadata
merged = []
for i, ch in enumerate(chunks_100):
merged.append({"id": f"100-{i}", "chunk_size": 100, "content": ch})
for i, ch in enumerate(chunks_400):
merged.append({"id": f"400-{i}", "chunk_size": 400, "content": ch})
# 5) Save merged chunks
with open(CHUNKS_MERGED_PATH, "w", encoding="utf-8") as f:
json.dump(merged, f, indent=2, ensure_ascii=False)
print(f"Saved {len(merged)} merged chunks -> {CHUNKS_MERGED_PATH}")
# 6) Build FAISS & BM25 on merged chunks
texts = [m["content"] for m in merged]
build_dense_faiss(texts, FAISS_PATH)
build_bm25(texts, BM25_PATH)
# 7) Save metadata
with open(META_PATH, "wb") as f:
pickle.dump(merged, f)
print(f"Saved metadata -> {META_PATH}")
print("\n✅ Done. Created 100 + 400 chunks separately and merged them for unified FAISS & BM25 indexes!")
if __name__ == "__main__":
main()