Rust RAG Engine

Most RAG demos stop at pure vector search. This project goes further: a two-stage hybrid retrieval pipeline where semantic pgvector cosine search narrows the candidate set, then a keyword re-rank pass scores within that set and fuses the two signals. The result handles exact-term queries that trip up pure vector search without sacrificing recall on paraphrased ones. The whole thing is written in Rust — not because Rust is fashionable, but because the architecture calls for it: an Actix actor model per WebSocket session, a bounded-concurrency Ollama pipeline, and a pgvector integration that needs zero-overhead async. The ingestion side fetches articles from a HelpScout-style REST API, converts HTML to clean Markdown using html2md and custom scraper rules, stores canonical articles in PostgreSQL via Diesel ORM, then fans out to two parallel pipelines: a Python Flask microservice (all-MiniLM-L6-v2) generates embeddings stored in pgvector, while a bounded-concurrency Ollama pipeline generates per-article keywords and summaries. The metadata pipeline uses a Tokio Semaphore and buffer_unordered to cap in-flight local LLM calls, with structured success/partial/failure accounting and failed-ID persistence for retries. Streaming chat runs through an Actix actor model. Each WebSocket connection spawns a ChatSession actor; a central ChatServer actor routes messages and manages session state. Ollama's streaming API is consumed token-by-token and forwarded over the connection in real time — the client sees words appear as they're generated, grounded in articles retrieved by the hybrid pipeline. A round-robin Ollama load balancer backed by a Rayon thread pool is included in the codebase for multi-instance inference distribution, currently shelved pending hardware but fully wired and ready. This is a backend-only service — the unfinished frontend scaffold was removed intentionally. The value is in the infrastructure: a production-grade Rust RAG pipeline with real hybrid retrieval, streaming chat, and a local-inference-first design. Clean cargo build, zero clippy warnings, 19 tests, CI configured. Public on GitHub under bredmond1019.