LLM-driven key-value store with confidence tracking

Look up any value by key. If it's not known, provide a prompt and the system figures it out. Confidence decays over time; re-assessment keeps values fresh.

A key-value store where values are determined by LLM assessment and improve over time. Each key has a history of observations with confidence scores. Reads walk the history newest-first and return the first assessment whose decayed confidence clears the caller's threshold. Request a minimum confidence — if nothing meets it, you get null.

Use this when

• You need a value that is uncertain or changes over time.
• You want to know something and have a prompt for how to figure it out.
• You need to track how confident you are in a piece of knowledge.
• You want to memoize LLM observations with decay.

Do not use this when

• You need to store operational data with a single authoritative value.
• You need a cache with TTL semantics.
• You need vector similarity search.
• You need to run inference directly (use complete).

Inputs

• A key (any string, often a hash).
• A prompt describing how to determine the value.
• A confidence threshold.
• Optional max_age duration to force a refresh.

Outputs

• A value with a confidence score, or null.
• Full assessment history with provenance.

Decision rule for AI agents

Choose lookup when you need a value that benefits from LLM assessment, may change over time, and should carry a confidence score. Think of it as memoized LLM inference with decay.

API endpoint

GET /api/lookup/:key