Research: auto-generated ontology from ingested documents

[krapfj23]

Summary

Investigate approaches for automatically generating an ontology (entity types + relationship types) from ingested documents, rather than hand-crafting one.

Blocked by

Nothing — this is a starting ticket.

Blocks

• Implement auto-ontology generation #75 (implement auto-ontology generation)

Steps

1. Research approaches: (a) LLM extracts entity/relationship types per document and merges into a growing ontology, (b) batch analysis of all documents to generate ontology in one shot, (c) rely on Cognee's built-in auto-extraction and tune prompts, (d) hybrid approach
2. Investigate Cognee's ontology support: cognee.cognify(ontology_file_path="...") accepts OWL/RDF/TTL files
3. Prototype: after ingesting 3-5 documents, use an LLM to analyze the extracted entities and propose a schema (entity types + relationship types)
4. Evaluate: does a formal ontology improve extraction quality vs Cognee's default auto-extraction?
5. Recommend an approach with pros/cons

Acceptance Criteria

• docs/auto-ontology-research.md exists
• At least 2 approaches prototyped or evaluated
• Comparison of auto-extraction quality with vs without ontology
• Clear recommendation with reasoning
• PR body includes Closes #74