Answers grounded in your documents. Every time.
A knowledge base isn't a file drawer — it's what keeps an AI honest. Akili indexes what you upload and answers with citations you can actually check.
Ingest, retrieve, cite
Ingest
Upload documents, paste URLs, or crawl a site. Content is chunked and embedded into a vector index automatically — with the indexing state visible, not hidden.
Retrieve
Hybrid retrieval fuses vector similarity with keyword search, expands to neighbouring passages, and re-ranks with time decay — so recent, relevant material wins.
Cite
Citations are enforced, not optional. Every grounded answer points back to the passages it used, so you check the claim instead of trusting the vibe.
Search that explains itself
Knowledge search returns passages, not just documents — with similarity scores explained honestly in the interface (cosine similarity, not a percentage, and not comparable across queries). If something isn't indexed yet, the UI says so instead of silently returning less.
Memory that compounds
The tenth conversation should be better than the first. Akili remembers what matters:
- Persistent facts about you that carry across sessions — preferences, context, terminology.
- Auto-extraction from conversations, plus import from your past chats.
- Admin governance for orgs: category restrictions and per-user caps, not a free-for-all.
- A library for account-scoped files — collections, share links, and team destinations.
FAQ
Knowledge and memory questions
Short answers, nothing hiding in a footnote.
Upload files directly — documents, spreadsheets, PDFs — paste URLs, or crawl a site to pull in a documentation set. You can also promote files from your library, the account-scoped file store, into any knowledge base.
No. Every query runs the same way: vector similarity and keyword search are fused, results expand to neighbouring passages for context, and re-ranking applies time decay so newer material outranks stale duplicates.
Grounded answers cite the passages they used. The correct way to evaluate this — or any — RAG system is to upload documents you know well, ask a question they answer, and check a citation. If a platform can't show you where an answer came from, it's asking you to trust the vibe.
They're cosine similarity between your query and the passage — not a percentage, and not comparable across different queries. The interface says so rather than dressing the number up as a confidence rating.
Indexing state is visible on every source. A knowledge base tells you when it's still processing rather than silently answering from half the corpus.
Free includes 3 knowledge bases with 50 documents each — enough to run a real evaluation. Pro includes 10, Max and Starter 25, and Team, Business, and Enterprise are unlimited. Team and Business also add shared, org-wide knowledge bases governed by roles.
Yes, on Business and Enterprise. Admins get category restrictions and per-user caps, so memory is governed rather than a free-for-all.
Point it at something you know well. Check the citations.
That's the whole evaluation. Free plan includes 3 knowledge bases with 50 documents each.