Talon weighs BM25-first semantic caching over embeddings

This is a sensible infra-first take on semantic caching: optimize for deterministic agent traffic before paying the complexity cost of embeddings.

• –BM25 fits Talon’s single-binary Go design and avoids bundling a local embedding model just to catch paraphrases.
• –For agentic workloads, retries and repeated task templates often matter more than natural-language variation, so exact or near-exact matching can go surprisingly far.
• –Optional embedding lookup through Ollama is a smart middle ground because it preserves local deployments without forcing extra dependencies on every user.
• –The false-hit concern is real: a bad semantic cache match in an LLM proxy can quietly serve the wrong answer, which is often worse than a clean miss.