Put all your PDFs and docs in, get cited charts, tables, and diagrams.
Keyline lets your agent make one call and retrieve precisely scoped evidence from a pre-indexed library of your docs.
Keyline API and CLI. Public Preview opens soon.Leave your email and we will send you a key and the docs the day it opens.
Contact salesRetrieve the original block that holds the answer.
Every page is analyzed into semantic blocks: text, figure, table, equation. A question resolves to the blocks that answer it, and each one comes back with the region and the page it was cut from.
Every page is read as an image.
A document carries meaning in more than its sentences, in tables, figures, diagrams, and equations. Reading each page as an image keeps all of it available, so an answer can come from a table as readily as from a paragraph.
Every block is described in words.
A figure or a table usually contains very little text of its own. At ingest a model writes a short description of what each block holds, and questions are matched against those descriptions.
Every answer carries its block.
A question resolves to a single region of a page. The answer arrives with that region attached, as a page number and an image of the block, so a reader can check the claim against the original.
Index once. Ask anything. Verify everything.
Index.
Send a document once. Keyline rasterizes every page, detects the semantic blocks on it, and builds a topic index you poll for readiness.
Ask.
One request, one natural-language question, the whole project as the corpus. Keyline searches, reads, and decides when it has enough.
Verify.
The answer arrives with numbered references, each resolving to a file, a page, and a cropped image of the block it came from.
What comes back
A streamed answer, a structured reference list, and the cropped region behind each claim. A citation that returns the actual figure from page 212 ends the conversation.
- Progress eventsThe retrieval loop narrating its own steps
- Answer textStreamed, with inline numbered citations
- Retrieved regionA PNG of the exact block that was cited
- Page referenceFile, page, and figure, resolved from the reference list
Request and response samples ship with Public Preview access.
What you can build
A retrieval primitive is most useful as a tool your agent reaches for.
A sub-agent for the questions that must be exact.
Your orchestrator hands off, and gets back an answer plus the page and the crop to show alongside it.
A field-service copilot over a legacy PDF library.
Service manuals, spec sheets, parts catalogs. The documents where the answer is a cell in a table.
A verification step in a workflow you already ship.
Before your pipeline acts on a dimension or a compatibility claim, ask for the source and attach it.
Anything a user will ask you to justify.
Regulated, safety-critical, or liability-bearing lookups. The crop is the audit trail, and you get it without building a viewer.
One layer with one job.
Keyline owns a narrow slice and hands the rest back to you.
| Keyline owns | How that works |
|---|---|
| Its own index | Built over the document library you send it, with no embeddings or chunking to configure |
| One retrieval question at a time | Called on request by your orchestrator, your model, and your framework |
| The documents your assistant struggles with | Scanned manuals, diagrams, compatibility tables |
| Text, page numbers, and images | Returned for your application to render in your own UI |
Server to server, with a bearer token from your own code, or from the Keyline CLI when the Public Preview opens. Pricing and published limits live on the pricing page.
Roadmap
Where every capability stands today. We would rather you find this than find a surprise.
| Agentic retrieval across a library | Available now |
| Deep read and vector search | Available now |
| Cited crops and page numbers | Available now |
| US region | Available now |
| Office file support | Available now |
| Self-serve signup and keys | Public Preview |
| Command-line tool | Public Preview |
| Public API documentation | Public Preview |
| Agent runbook and skill | Public Preview |
| Region coordinates for cited blocks | Public Preview |
| Webhooks | After preview |
| Rate limiting and quotas | After preview |
Why this exists
Keyline was built for manufacturers whose answers live inside thousands of legacy PDFs and who needed to verify every one of them.
At one equipment manufacturer, a single engineer was the answer desk for 218 service engineers across 70 locations. After 370 pages were made searchable, monthly access rose around five times and inquiry handling workload fell roughly 80 percent.
That is where the retrieval behavior comes from, and why accuracy is the axis it was tuned on. Keyline is funded and proven at Helpfeel, where 625 companies run our customer support platform. The API is the same engine, exposed.
Sign up for the Public Preview
Leave your email and tell us what you are building. You get a key, $50 in credit, and the documentation the day it opens.
Contact salesBerkeley, SF, and anywhere else: if you are at an event with us, say hello and we will put you at the front of the list.
FAQ
What a citation contains, how to tell when a document is ready, which files are supported, what a key costs, and how to get one. The answers live in the documentation.
Learn more in the docsKeyline is the document intelligence product from Helpfeel. See the pricing and limits, the agent runbook, the product overview, or the Helpfeel platform.