Semantic Retrieval Tuning and Value-First Docs

Boosting Semantic Recall

Hybrid search typically relies on Reciprocal Rank Fusion (RRF) to combine signals from exact-key matches, full-text search (FTS), and vector KNN. While RRF is excellent for merging multiple signals, it inherently penalizes single-signal results. This created a discovery issue: if a user searched for “siblings” but the stored knowledge used the word “sister,” the FTS signal would return zero results, leaving only the vector signal.

Because RRF scores for single signals are mathematically much lower than multi-signal hits, these semantic matches were often buried. I’ve implemented a 2.5x boost for vector-only scores specifically when both exact-key and FTS signals return empty. This ensures that semantic-only hits surface to the user when no literal word overlap exists.

Refined Project Positioning

The documentation received a significant overhaul today. I moved away from architecture-heavy descriptions and generic AI imagery in favor of a technical, value-driven README.

The new messaging centers on Chalie as a “persistent reasoning engine” rather than just a chatbot wrapper. The focus is now on three core pillars: compounding memory, autonomous goal formation, and background cognition (thinking while you’re not looking). I also simplified the installation and local-first setup instructions to get developers running with Ollama more quickly.