feat(ingestion): document ingestion — upload doc, Ei extracts, persona surfaces #37
Description
Summary
A workflow for uploading a document (PDF, DOCX, plain text), having Ei break it down into facts/topics/entities, and then routing that knowledge to a persona specifically configured to surface and discuss it.
Use Case
Flare mentioned co-workers who might benefit from: upload a spec doc → Ei extracts relevant facts and topics → a "Spec Reviewer" persona can answer questions about it, surface relevant sections, and connect it to other things Ei knows.
Distinction from Existing Behavior
Current extraction pipeline:
- Triggered by conversations (user sends a message → Ei extracts from the exchange)
- Sources: chat, OpenCode sessions, Claude Code sessions, Cursor sessions
Document ingestion:
- Triggered by a file (not a conversation)
- Sources: PDF, DOCX, Markdown, plain text
- Extraction would run directly on document content, not conversation turns
Open Questions
- Does extracted knowledge go to the human entity (global) or stay scoped to a specific persona?
- How is authorship/source tracked? (The facts came from "Q3 Roadmap.pdf", not from a conversation)
- Does the document get chunked? (Large docs won't fit in a single extraction context)
- UI: how does a user upload in web vs TUI?
Notes
Not near-term — planting the flag. When prioritized, revisit in context of the extraction pipeline architecture.