Biblicus

Make your documents usable by your assistant, then decide later how you will search and retrieve them.

Taxonomy and ontology

Taxonomy and ontology are the reviewed structure that turns topic discovery into graph steering.

Mental model

The taxonomy is a strict topic tree. Root nodes usually come from accepted canonical topics. Child nodes describe reviewed subtopics under those roots. Each taxonomy node has a stable topic_uid, a display name, a description, an optional parent, and reviewed seed or holdout evidence.

The ontology is a graph of typed relationship assertions. Relationship types define domain vocabulary such as influenced, historical_context_for, or use_case_of. Assertions connect item:<item_id>, topic:<topic_uid>, and graph:<node_id> references.

Neo4j is a materialized GraphRAG graph. It can include graph extraction output, taxonomy topic nodes, taxonomy subtopic_of edges, item member_of_topic edges, and accepted ontology assertions.

Accepted taxonomy input

The steering application exports accepted taxonomy JSON. Biblicus validates and records it:

biblicus taxonomy record \
  --corpus corpora/example \
  --input accepted-taxonomy.json

Taxonomy input shape:

{
  "schema_version": 1,
  "taxonomy_id": "example-taxonomy",
  "display_name": "Example Taxonomy",
  "description": "Accepted topic hierarchy.",
  "generated_at": "2026-05-16T00:00:00+00:00",
  "nodes": [
    {
      "topic_uid": "agent-systems",
      "parent_topic_uid": null,
      "display_name": "Agent Systems",
      "description": "Research about agent architectures and behaviors.",
      "status": "accepted",
      "seed_item_ids": ["item-one"],
      "holdout_item_ids": []
    },
    {
      "topic_uid": "agent-memory",
      "parent_topic_uid": "agent-systems",
      "display_name": "Agent Memory",
      "description": "Research about memory in agent systems.",
      "status": "accepted",
      "seed_item_ids": ["item-two"],
      "holdout_item_ids": []
    }
  ]
}

Biblicus rejects duplicate topic identities, unknown parents, cycles, and unknown fields.

Taxonomy discovery

Discovery finds possible child nodes under accepted roots:

biblicus taxonomy discover \
  --corpus corpora/example \
  --classifier example-classifier \
  --extraction-snapshot pipeline:<snapshot_id> \
  --steering-feedback papyrus-steering-feedback.json \
  --format markdown

Discovery uses classifier topic membership to collect documents under each accepted root, runs scoped topic modeling for that root, and emits steering proposals such as create-taxonomy-node. It does not mutate the accepted taxonomy. When --steering-feedback is provided, Biblicus validates the Papyrus feedback export and suppresses rejected child-topic patterns that match the same classifier and root topic. Suppressed candidates are reported as warnings rather than written as new proposal records.

Accepted ontology input

The steering application exports accepted relationship types and assertions. Biblicus validates and records them:

biblicus ontology record \
  --corpus corpora/example \
  --input accepted-ontology.json

Ontology input shape:

{
  "schema_version": 1,
  "ontology_id": "example-ontology",
  "display_name": "Example Ontology",
  "description": "Accepted domain relationships.",
  "generated_at": "2026-05-16T00:00:00+00:00",
  "relationship_types": [
    {
      "relationship_uid": "influenced",
      "display_name": "Influenced",
      "description": "The source influenced development of the target.",
      "directed": true
    }
  ],
  "assertions": [
    {
      "assertion_id": "assertion-one",
      "source_ref": "item:item-one",
      "relationship_uid": "influenced",
      "target_ref": "topic:agent-systems",
      "direction": "outbound",
      "evidence_item_ids": ["item-one"],
      "confidence": 0.8
    }
  ]
}

Graph materialization

Materialization is explicit:

biblicus ontology apply \
  --corpus corpora/example \
  --taxonomy <taxonomy_snapshot_id> \
  --relationships <ontology_snapshot_id> \
  --graph-snapshot simple-entities:<graph_snapshot_id>

The command writes deterministic topic nodes, subtopic_of edges, item member_of_topic edges, and accepted ontology relationship edges into Neo4j as an ontology overlay for the selected graph snapshot.

Query accepted ontology assertions:

biblicus ontology query \
  --corpus corpora/example \
  --relationships <ontology_snapshot_id> \
  --source-ref item:item-one \
  --relationship influenced \
  --direction outbound