Execute
Python JavaScript Bash and SQL run in a separate locked executor container.
CortexLab V8
From local intelligence features to local intelligence infrastructure.
V8 separates execution retrieval reranking tools evaluation and language generation into inspectable components. The goal is not to imitate a frontier cloud model. The goal is to own the stack and measure every layer.
Index
A persistent graph ANN artifact narrows semantic candidates before reranking.
Rerank
A trainable relevance model combines lexical semantic graph ANN exact scope and priority signals.
Orchestrate
Agents receive tools from an explicit registry with risk classes and hard timeouts.
Evaluate
Benchmarks track intent topic MRR recall and retrieval regressions.
Generate
An internal llama.cpp service can host a local GGUF small language model.
agent tool registry
Autonomy only through named tools.
read
hybrid.retrieve
Retrieve grounded CortexLab evidence
1800 ms timeoutcompute
ann.search
Search persistent-style ANN knowledge graph
1000 ms timeoutread
knowledge.lookup
Look up exact local concepts
600 ms timeout