Forge Capsule

Vector Databases and Embedding Search at Scale

Vector search fundamentals: approximate nearest neighbor (ANN) search, HNSW (Hierarchical Navigable Small World) index, IVF (Inverted File Index). Faiss: flat L2/cosine, IVFFlat, IVFPQ, HNSW. Production vector DBs: Pinecone (managed, serverless), Weaviate (multi-modal, hybrid BM25+vector), Qdrant (Rust, payload filtering), Milvus (distributed, GPU). Embedding models: OpenAI text-embedding-3, Cohere embed-v3, BGE-M3 (multi-lingual, multi-granularity). Hybrid search: RRF (Reciprocal Rank Fusion) merges lexical + semantic scores. Chunking: semantic (sentence-transformer boundaries), recursive character, hierarchical. Forge uses hybrid retrieval combining vector cosine similarity with Neo4j graph traversal.