.latch
Compact, privacy-first bundle. Carries the compiled latent state but omits extracted source text. Use it when the destination only needs the reasoning surface and should not carry readable documents.
Bundles
.latch and .latchdoc are the portable bundle formats for LATCH. They let one operator pay the document-processing cost once, then move the finished compiled corpus to another compatible runtime without recompiling, re-extracting, or re-reading the original source set.
Format Thesis
PDF made document layout portable. LATCH bundles do the AI-era equivalent for compiled document intelligence. The expensive steps such as extraction, tokenization, and neural compilation are paid once by the creator. Every recipient opens the finished result and starts querying immediately.
Compact, privacy-first bundle. Carries the compiled latent state but omits extracted source text. Use it when the destination only needs the reasoning surface and should not carry readable documents.
Full-fidelity bundle. Carries compiled latent state plus extracted source text for full-text lookup and text-based fallback. This is the recommended default for most operator-facing workflows.
Use Cases
Compile the diligence corpus once, share the finished file with multiple analysts, and keep exact source language available for memos, committee materials, and fallback review.
Move the compiled reasoning surface across underwriting rooms without shipping readable policyholder text, loss runs, or other downstream-sensitive source material.
Use when compliance teams need cross-document reasoning plus exact text lookup for audit responses, policy review, and citation-heavy work.
Ship compiled world knowledge as a runtime asset. The destination machine gets fast, portable memory without carrying a human-readable lore corpus.
FAQ
LATCH is a persistent document-intelligence runtime that compiles a document set into reusable latent memory instead of re-reading the raw corpus on every query.
RAG retrieves chunks per question. LATCH compiles the corpus into a persistent representation so cross-document reasoning does not depend on chunk retrieval quality at query time.
Large-context prompting pays the full document-read cost on every query. LATCH pays the compilation cost once, then amortizes it across all later queries and recipients of the bundle.
The runtime and bundles stay local. The only network dependency in the supported customer path is the startup license validation call.
The current validated lineup is centered on Qwen 2.5 14B, with additional supported tuples across the Llama, Mistral, and DeepSeek families on the product roadmap and internal validation path.
FAQ
Use .latch when the destination should get the compiled reasoning surface but should not carry readable extracted text. This is the privacy-first choice.
Use .latchdoc when the destination also needs exact source language, full-text lookup, and text-based fallback for edge cases.
Fallback is the quality safety net for .latchdoc. If the primary LATCH answer is refusal-like, low-confidence, or errors, the runtime can fall back to the embedded text path instead of leaving the query stranded.
No. Fallback needs embedded source text, so it is available only on .latchdoc. That tradeoff is intentional.
Size depends on corpus size and memory settings, but typical enterprise bundles land in the low hundreds of megabytes. .latchdoc usually adds only modest text overhead compared with the compiled tensor payload.
FAQ
The recipient avoids re-extraction, re-tokenization, and recompilation. They inherit the finished compiled artifact instead of paying the ingestion pipeline again.
Yes. LATCH bundles are encrypted and portable only across compatible runtimes. Team sharing can use the instance license token or an explicit passphrase when cross-license sharing is required.
The supported privacy posture is that .latch omits extracted source text and carries compiled latent state rather than a readable copy of the original documents.
The primary validated path targets NVIDIA H100 80GB and A100 80GB class rooms. Customer-facing quickstarts currently assume that class of GPU.
FAQ
LATCH ships as a Docker image. The supported quickstarts cover private ACR pull, startup token placement, and first-room validation for RunPod and Ubuntu GPU hosts.
The current product path supports PDF, TXT, MD, HTML, DOCX, XLSX, PPTX, CSV, JSON, and XML.
The public evaluation path includes the self-hosted runtime, Console UI, local API, and current v1 update stream under the applicable customer license terms.
Production deployment and embedded distribution remain direct-conversation paths because the right terms depend on the workload, distribution model, and support expectations.
The supported local runtime exposes an OpenAI-format API surface, so many existing tools and internal integrations can switch to LATCH with an endpoint change rather than a full client rewrite.