Part 5 · Getting Reliable JSON Out of a Local LLM
format="json" gets you 90% of the way there. The other 10% is trailing commas, shape surprises, and items that almost-but-don't-quite match the schema. Here's the defensive pattern …
Part 4 · Crash-Resumable Ingestion: DBOS, SHA-256, and Surviving a kill -9
Re-embedding a 200-page PDF every time you tweak one paragraph is a tax nobody wants to pay. Here's how CogniVault uses DBOS workflows and content hashing to ingest only what …
Part 3 · Two-Phase Streaming: Showing the Model Think Before It Acts
Most agent UIs hide the model's reasoning until everything finishes. CogniVault streams Gemma 4's chain of thought first, then runs the Strands tool loop — and the UX dividends are …
Part 2 · Hybrid Retrieval in Practice: FAISS + BM25, Fused with RRF
Dense vectors are smart but forgetful. Keyword search is dumb but loyal. Here's how I combined FAISS, BM25, and Reciprocal Rank Fusion in CogniVault — and why pure semantic search …
Part 1 · Why I Built a Local-First RAG
Cloud AI assistants are powerful — but for trainers, researchers, and anyone handling sensitive material, they're also a leaky abstraction. Here's why I built a 100% local …