Hi Product Hunt, I'm Kathiravan, the maker of Sulci.

Sulci is a semantic cache: when a new question means the same thing as one your app already answered, it returns the cached answer instead of calling the model again. I built it after counting what an agent actually sends to its LLM. In our benchmark the median 200-dispatch agent session sent 2 of them to the model once the cache was warm.

What I'm proudest of is a number most caches don't publish: on a corpus where 39% of questions have no correct cached answer, Sulci correctly says "I don’t have that" 98.9% of the time (98.15–99.38% across four seeded draws).

It's context-aware: give it a session and lookups take the conversation into account. On the hosted tier, an answer that only makes sense inside one conversation is never shared with another.

Adopting it costs a line or less: one line for LangChain/LangGraph, one call for LiteLLM, a base-URL change for any OpenAI or Anthropic client, or an MCP server for agents that can't import Python. It's open source (Apache 2.0), runs on SQLite, Chroma, Qdrant, FAISS, Redis or Milvus, and the hosted tier and its free dashboard went generally available this week.

I'd love feedback, especially if you try it on your own traffic and the numbers disagree with mine.

Hi Product Hunt, I'm Kathiravan, the maker of Sulci.

It's context-aware: give it a session and lookups take the conversation into account. On the hosted tier, an answer that only makes sense inside one conversation is never shared with another.

I'd love feedback, especially if you try it on your own traffic and the numbers disagree with mine.

Sulci

Semantic cache for LLM apps and AI agents

Semantic cache for LLM apps and AI agents