[feat]: Seed document

[invidtiv]

Problem statement

The ingestion of the data and the creation of the wiki sometimes fails due to lack of knowledge. There’s a lack of context when creating the embeddings and when extracting the information for the wiki.

Proposed solution

From what I’ve seen, this could improve by having a seed document for each chat. For instance, for the company, it would be the definition protocols and a definition of who is who.

That way, the wiki would produce better results and more real concepts. I’ve been modifying it, and so far what I did in my version was exactly that: creating the notion of the seed document and having it ingest the seed document when doing both tasks:

• creating the wiki
• ingesting the data

This greatly improved the outcome. So perhaps an onboard procedure where we can define the background context and vocabulary of a group, and I would also propose having a bigger classification, I use family, work, projects, and this makes also a group of chats collaborate into the same wiki , and use the same seed document.

I use AI to code, so not sure if I should do a PR.

Alternatives considered

No response

Additional context

No response

[jhkchan]

Thanks for the detailed write-up — you've identified a real gap. Heads-up: this mechanism already exists in demo/, so you can probably skip building it from scratch and instead enrich what's there.

What's already in place

• demo/corpus/*.md — seed documents (7 Wikipedia-derived markdown files about Ada Lovelace, Babbage, Python history)
• demo/seed.py --live — pumps every file through the same ADK ingestion pipeline as live chat (fact_extractor → entity_extractor → persister)
• Result: facts + entities land in Mongo/Weaviate/Neo4j keyed to a channel ID, so subsequent message ingestion inherits those priors via entity-similarity matching, and wiki compilation pulls from the enriched stores

That's exactly the "seed document for each chat → ingest before/alongside live data" flow you described.

Where a PR adds the most value

demo/seed.py currently hardcodes the channel + corpus path. The smallest mergeable enrichment is to generalize it:

• --channel-id <id> (required)
• --corpus-dir <path> (default: demo/corpus)
• Per-(channel_id, corpus-hash) idempotency so re-running with edited seeds re-ingests cleanly

Bonus: that incidentally covers your "family / work / projects" use case — run the script once per channel with the relevant corpus. AI-assisted PR is fine; just keep this scope tight.

Out of scope here

The "multiple chats share one wiki" piece is a separate architectural change (data model is currently channel-centric, no WikiBundle). Worth a dedicated issue once the seed-doc work lands.