Text graphs and topics
Historia text graphs are rebuildable structural projections of immutable message revisions. Every evidence-bearing node and edge points back to an exact UTF-8 byte range in a typed message block.
The first analyzer is deterministic and non-neural. It extracts source structure, configured project entities, explicit artifacts, shallow propositions, and rule-derived speech acts such as requests, constraints, proposals, decisions, acceptance, rejection, and status.
A graph-backed topic projection then turns reference nodes, proposition phrases, and keywords into retrieval handles. Work and discourse nodes remain graph structure: they qualify nearby topics and improve ranking instead of becoming unsupported semantic claims.
Build missing graphs and topics after importing conversations:
historia graph indexInspect the complete graph or one reversible projection:
historia graph show <revision-oid>historia graph show <revision-oid> --projection sourcehistoria graph show <revision-oid> --projection conceptshistoria graph show <revision-oid> --projection workInspect the topic plan for a query:
historia topic search "private agent rooms"Use the topic graph to expand ordinary lexical retrieval:
historia chat search "private agent communication" --expand-topics
historia context build \ "private agent communication" \ --expand-topics \ --budget 12000The topic layer stores direct graph-node links, local proposition context, work/discourse facets, and bounded evidence for one-hop associations. Context bundles distinguish lexical matches, direct-topic matches, and associated-topic matches so an LLM can understand why each original message was selected.
Git remains authoritative. SQLite stores the complete derived graph, topic mentions, and topic associations; all can be deleted and rebuilt.
The message-local analyzer still does not infer cross-message acceptance, rejection, answers, or task transitions. Those relations require a later conversation-snapshot projector over Historia’s reply graph, branches, and active paths.
See the text graph specification and topic index specification.