Part 8 · Testing a Local-AI App: 351 Tests, Zero Infrastructure
Tests that need Ollama running, or Postgres up, or a real PDF on disk, will be skipped in CI and rotted in a week. Here's how I made CogniVault's test suite run in any environment …
Gemma CogniVault
A fully local, privacy-first AI Study Companion — Gemma 4 + FAISS + BM25 + Strands Agents, running entirely on your machine.
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 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 …