Refactor embedding model integration to use fastembed and update related dependencies

build_faiss_index.py

@@ -23,7 +23,7 @@ import pickle
 import time
 import os
 from pathlib import Path
-from sentence_transformers import SentenceTransformer
+from fastembed import TextEmbedding
 
 # Paths
 SCRIPT_DIR = Path(__file__).parent.absolute()
@@ -43,20 +43,11 @@ def build_faiss_index():
         print("   Run ingest_data.py first to create the database.")
         return False
 
-    # Load sentence transformer model
-    print("\n1. Loading sentence transformer model...")
+    # Load embedding model
+    print("\n1. Loading embedding model (fastembed/ONNX)...")
     start = time.time()
 
-    # Disable all parallelism to avoid Python 3.14 issues
-    os.environ['TOKENIZERS_PARALLELISM'] = 'false'
-    os.environ['OMP_NUM_THREADS'] = '1'
-    os.environ['MKL_NUM_THREADS'] = '1'
-    os.environ['OPENBLAS_NUM_THREADS'] = '1'
-
-    import torch
-    torch.set_num_threads(1)
-
-    model = SentenceTransformer('all-MiniLM-L6-v2')
+    model = TextEmbedding('sentence-transformers/all-MiniLM-L6-v2')
     print(f"   ✓ Model loaded in {time.time() - start:.3f}s")
 
     # Load biographies from database
@@ -97,13 +88,7 @@ def build_faiss_index():
 
     for i in range(0, len(texts), batch_size):
         batch = texts[i:i + batch_size]
-        batch_embeddings = model.encode(
-            batch,
-            show_progress_bar=False,
-            convert_to_numpy=True,
-            normalize_embeddings=False,
-            device='cpu'  # Explicit CPU to avoid issues
-        )
+        batch_embeddings = list(model.embed(batch))
         embeddings.extend(batch_embeddings)
 
         # Progress update every 100 batches (~3200 texts)

congress_bio_ids.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a8b5e408f8c8522843b8dbe7cc8141de36ef80dbbd73495c0efdc3ed443b8d2
+oid sha256:a2df06881ad172440055bfdd84607704a18d0f1c05f2929bb9481ab5b0d9f2aa
 size 130521

congress_faiss.index

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e1a1761d05411f7b501a70da6b561724d4ca1d25e21a818747efa6e1b3114444
+oid sha256:b02b6f4ca7b3d8a9a59d42795d875badcea22b40031e03d8acc5dff21467fb08
 size 20040237

gradio_app.py

@@ -11,7 +11,7 @@ import os
 import warnings
 from typing import List, Dict, Any
 import numpy as np
-from sentence_transformers import SentenceTransformer
+from fastembed import TextEmbedding
 import faiss
 import pickle
 from pathlib import Path
@@ -39,7 +39,7 @@ def initialize_search_index():
     try:
         if Path(FAISS_INDEX_PATH).exists() and Path(BIO_IDS_PATH).exists():
             print(f"Loading FAISS index from: {FAISS_INDEX_PATH}")
-            model = SentenceTransformer('all-MiniLM-L6-v2')
+            model = TextEmbedding('sentence-transformers/all-MiniLM-L6-v2')
             faiss_index = faiss.read_index(FAISS_INDEX_PATH)
             with open(BIO_IDS_PATH, "rb") as f:
                 bio_id_mapping = pickle.load(f)
@@ -227,8 +227,7 @@ def semantic_search_biography(query: str, top_k: int = 5) -> str:
         top_k = min(max(1, top_k), 20)
 
         # Encode query
-        query_embedding = model.encode([query])[0].astype('float32')
-        query_embedding = query_embedding.reshape(1, -1)
+        query_embedding = np.array(next(model.embed([query])), dtype='float32').reshape(1, -1)
         faiss.normalize_L2(query_embedding)
 
         # Search

requirements.txt

@@ -1,7 +1,5 @@
-# Requires Python 3.10-3.13 (NOT 3.14+ due to FAISS incompatibility)
 mcp>=1.0.0
 numpy>=1.24.0
-sentence-transformers>=2.2.0
-torch>=2.0.0
+fastembed>=0.3.0
 faiss-cpu>=1.7.4
 gradio>=5.0.0