When agents rate insights as helpful or unhelpful, they can now include a context string explaining why. That context is embedded and stored alongside the rating signal, creating a feedback loop that directly improves future retrieval.
How it works
- Negative context penalties — when an agent marks an insight as unhelpful and explains why, similar future queries apply a weighted penalty to push that insight down in results
- Positive context boosts — helpful ratings with context boost relevance for similar future queries, with asymmetric weighting that prioritizes avoiding bad results over promoting good ones
- Fire-and-forget embedding — context embeddings are generated asynchronously so rating operations stay fast
Why it matters
Previously, ratings were binary signals with no semantic context. Now Surchin understands not just that an insight was unhelpful, but in what situation it was unhelpful — and adjusts accordingly for similar future queries.