Index once, reuse forever: durable long-context AI

Most AI stacks re-read your documents from scratch on every call. lab358 indexes a document once and reuses that context across every chat, agent, and API request — no repeated prefill, no fixed context ceiling.

Ask a question about a 300-page contract today and the model reads all 300 pages before it answers. Ask a second question an hour later and it reads all 300 pages again. Point an agent at the same repository across a dozen tool calls and it re-reads the repository a dozen times. In a standard stack, the context is thrown away the instant the response is returned — so the most expensive part of the request, processing the document, is paid for over and over.

That is backwards. The document didn't change between questions. The work of understanding it shouldn't be repeated.

Index once, reuse everywhere

lab358 treats a document's context as a durable, reusable asset. You index a document once; its context is computed a single time and then reused on every later request against it — the next question, tomorrow's question, a different chat, a different agent, a batch job. There is no prefill to redo between uses, even days apart.

The result is a different mental model for building with long context:

• Chats over the same corpus stop paying the re-reading tax on every turn.
• Agents that loop over a fixed set of documents index them once and reuse that context across every step, instead of re-ingesting the corpus each tool call.
• API calls against a shared corpus — a whole team, a whole app — reuse the same indexed context, so the savings compound the more the corpus is hit.

How it works, briefly

lab358 converts the open-weight model you already use into a natively sparse, top-k attention architecture, and serves it with approximate nearest-neighbor retrieval over the model's key–value cache. Two things follow from that:

• The context is computed once and cached. Because retrieval reads from a stored cache rather than re-processing tokens, a document's cache becomes a reusable asset — index once, reuse on every request.
• The ceiling comes off. The cache no longer has to live in GPU memory, so context length isn't bounded by a single device — the long documents that standard attention punishes are exactly the ones this unlocks: contracts, case law, patient histories, filings, entire codebases.

And nothing about your stack changes to get it: the converted model exposes an OpenAI-compatible API, so it's a base-URL change, not a rewrite.

Availability

lab358 is coming two ways — the same converted model and the same OpenAI-compatible API either way, so you pick where it runs:

• lab358 Cloud. A fully managed, usage-based workspace hosted by lab358 — nothing to operate. It's rolling out gradually: we're in staging testing now, then early access with a small group of trusted users, then opening by waitlist. Join the waitlist and we'll reach out as access expands.
• Self-hosted. The same converted model as an AWS Marketplace container product — a Helm chart that deploys into your own Amazon EKS cluster, running entirely inside your AWS account, for teams that need it in their own perimeter.

If you build on long-context documents, join the lab358 Cloud waitlist — or, to talk about self-hosting or getting your model family on the supported list, talk to us.

Investors: hello@lab358.ai.