LinkingMem - v0.3.0

Every RAG system out there is running two separate databases—some vector store like Pinecone or Qdrant for embeddings, Neo4j or similar for the knowledge graph—and then duct-taping them together with fragile sync pipelines. That's a $10 billion industry held together with bubble gum and prayer. LinkingMem said no.

LinkingMem is a graph-native RAG engine built in Rust and Python that merges vector search, graph traversal, and LLM reasoning into ONE pipeline. Query flows through Embed → HNSW → BFS → Score → LLM answer. The CSR graph and HNSW vector index share memory, no separate databases, no sync nightmares. It handles text AND images in the same vector space—you can query by text, by image URL, or both, with caption or CLIP embedding backends. And the multi-hop reasoning lets the LLM iteratively request graph expansion via /query/multihop before generating final answers.

The real flex? It works with ANY OpenAI-compatible provider—OpenAI, Ollama, Gemini, Groq, LM Studio, vLLM. Just point OPENAI_BASE_URL. It's production-ready with delta stores, WAL crash recovery, hot-swap graph merging, Prometheus metrics, per-key rate limiting, and distributed ingest. There's even a plugin interface so you can swap in custom embed/extract/generate servers in any language over plain HTTP or Unix sockets.

Most RAG startups are just wrapping APIs. LinkingMem actually rebuilt the engine. This is infrastructure code, not wrapper code. If you're building AI products and tired of managing three different databases just to get decent retrieval, this is the shortcut.