Skip to content
Historia

Begin typing to search this documentation.

Neural topic retrieval

Historia’s neural layer is an optional ranking projection over the existing text graph and topic index. It encodes immutable message revisions and topic labels, then retrieves semantically nearby topics and their supporting original messages.

It does not replace lexical search or make semantic claims. Every result still includes the source ref, archive commit, message revision, graph evidence, and original text.

Pipeline

Section titled “Pipeline”
immutable message revision
  → anchored text graph
  → graph-backed topic mentions
  → optional MiniLM vectors
  → direct and one-hop topic ranking
  → original messages for an LLM

Commands

Section titled “Commands”
Terminal window
bun add @huggingface/transformers
historia neural index
historia neural search "private channels for external agents"
historia neural status

CPU/WASM defaults to a quantized profile. WebGPU can use fp16 where supported:

Terminal window
historia neural index --device webgpu --dtype fp16 --batch-size 32

The model is pinned by revision and cached by a fingerprint that includes its runtime, device, dtype, pooling, normalization, dimensions, and classifier prototypes. The neural tables are rebuildable SQLite projections; Git remains authoritative.

See spec/chat-neural-index.md in the source repository for the storage, classification, retrieval, and privacy contracts.