Documentation
¶
Overview ¶
Command judgment-semantic-quickstart wires semantic judgment memory over a local OpenAI-compatible embedding server.
Tier 0: a quantized local embedding model (default nomicai-modernbert-embed-base-4bit) behind an OpenAI-compatible /v1/embeddings endpoint embeds every judgment input. Tier 1: a reflex Jev judge answers the typed primitives. Tier 2: below-threshold reflex confidence escalates to a deliberative Jev judge. The judgment/memory SemanticStore caches served judgments keyed by their input embedding, so a later paraphrase of the same question is answered from memory — same judgment, zero provider calls — as long as its embedding stays within the similarity threshold.
The judgment.Memory router contract carries no context; this example owns one context from run() and reuses it for the store's embed calls. Both Jev providers are deterministic scripted stand-ins; real deployments inject a jev.Client that owns transport and credentials.
The caller owns the embedding server process. Start it first (defaults target http://localhost:8000/v1), then run from the repo root:
GOWORK=off go run ./examples/judgment-semantic-quickstart
Override with EMBED_SERVER_URL, EMBED_SERVER_API_KEY, EMBED_MODEL, EMBED_DIM, and EMBED_THRESHOLD.
Source Files