Cite the source.
Documents in. Ranked passages out — attached, not recalled.
Ingest. Index. Retrieve — with receipts.
Ingest
Upload files or point at a site. Extraction handles PDFs, office documents, and pages — including files already in the agent's Workspace.
Index
Content is chunked and embedded for relevance-ranked search, billed per million tokens processed — so re-index only what changes.
Retrieve
At run time the agent pulls the most relevant passages into context and cites them in its answer. Every search is journaled, like any other capability call, and billed at its published per-query rate.
Answers with the source attached.
knowledge_search("termination notice period — Acme MSA") → 3 passages · 214ms [1] acme-msa-2025.pdf · §12.4 "…either party may terminate with sixty (60) days' written notice…" [2] acme-msa-2025.pdf · §12.6 "…obligations surviving termination…" [3] renewal-addendum.pdf · §2 "…auto-renewal unless notice per §12.4…" agent: Sixty days' written notice, per §12.4 of the MSA [1] — and note the addendum auto-renews unless that notice lands [3].
Knowledge is what you give the agent; memory is what it learns on the job. They work together — retrieval grounds the facts, memory keeps the relationship — but they're separate stores with separate rules.
Three published rates. Nothing hidden.
| Meter | Unit | Price |
|---|---|---|
| Indexed storage | per GB / month (vector + search structures) | $1.50 |
| Knowledge search | per query | $0.00002 |
| Knowledge indexing | per 1M tokens processed | $0.20 |
| Embedding models | provider price + $0.25 / 1M tokens | Metered |
Your documents stay yours.
Every document, chunk, and vector is isolated to your account. Retrieval can't cross tenants.
Every retrieval is on the record — what was asked, what was returned, and when.
Ingested content serves your agents. It doesn't train models — ours or anyone's.
Ground an agent in your documents.
Ingest something real and ask about it — the $10 credit covers a lot of retrievals.