Ommatidia

What it knows

The literature a web search is worst at.

Defense, Department of Energy and NASA technical reports: decades of physics, materials, propulsion and astronomy work whose full text sits in archives a consumer search engine barely touches. Not abstracts — the whole document, split into passages, each citable by a stable handle your assistant can quote back.

DTIC3.6M passages — aerodynamics, structures, propulsion, materials under load. Defense technical reports going back to the 1950s.

OSTI2.8M passages — condensed matter, nuclear and energy systems, plasma, national-laboratory research.

NASA ADS2.8M passages — observational and theoretical astronomy, astrophysics and planetary science.

NASA NTRSFlight test, spacecraft structures, thermal protection and mission engineering.

Choose what you install

Take the archives you need, not all of them.

Each archive is a separate download. An astronomer has no use for 45 GB of defense structures reports, and a 16 GB laptop should not be told it needs a workstation. Pick below and the requirements update.

DTIC Aerodynamics, structures, propulsion, materials under load. Defense technical reports from the 1950s on. 3.6M passages
66.7 GB OSTI Condensed matter, nuclear and energy systems, plasma, national-laboratory research. 2.8M passages
51.7 GB NASA ADS Observational and theoretical astronomy, astrophysics, planetary science. 2.8M passages
52.6 GB NASA NTRS Flight test, spacecraft structures, thermal protection, mission engineering. 5k passages
0.1 GB

Low-memory mode Keeps the search graph on disk instead of in RAM. Cuts memory by about two thirds and costs roughly 3× on query latency — a fraction of a second becomes about a second. The right choice on 16 GB. Ship a language model too (Gemma 4 E4B) Off by default: your agent already has a model. Turn it on for a stack with no outside model at all. Needs an NVIDIA GPU with 12 GB or more. Figures (paid) Plots, spectra and micrographs matched to the passage that cites them. Downloads once, then fully offline.

This machine needs

Disk: —
Memory: —
GPU: —
Passages: —

How the privacy works

The query never leaves. Structurally, not as a promise.

Your assistant talks to a container on loopback. That container reads an index and a database on your disk. Nothing in that path has a route to the internet, which is why you can test the claim by unplugging the machine.

Every arrow inside the boundary is a local socket or a file read. The single dashed crossing is the corpus download, and it happens once. After it finishes you can disconnect permanently and searching still works.

What costs money

The text is free. The plots are not.

The literature is public record, so it is free and stays free. What we charge for is the extracted figure corpus — the wind-tunnel plots, stress curves, spectra and micrographs pulled out of those reports and matched to the passage that cites them — and, later, licensed proprietary collections. In physical sciences the figure often is the result.

Ommatidia

Free

9.2M passages, full text
77.1M vector index
MCP endpoint for any agent
Signed corpus updates
Works with the NIC unplugged
No figures

Figures

Paid, one-time

10.4M plots, spectra and diagrams with their original captions
Matched to the passage that cites them
Downloads once, then fully offline
No per-query call, ever
Unlocks in place: ommatidia unlock

Upgrading does not change the privacy posture, and that is a design constraint rather than a marketing line. A paid bundle is bought once and decrypted locally, so we learn that a key was redeemed and nothing after that — not what you searched, not what you opened, not whether you are still running it. The alternative design, metering each figure as you view it, is what our enterprise deployments use because publisher contracts require per-access accounting. It is the wrong trade for a personal tool, so it is not the default here.

Getting it running

One command, then tell your agent it exists.

The installer checks disk, memory and ports before it downloads anything, then registers itself with whatever agent it finds.

$ curl -fsSL https://ommatidia.ai/install.sh | bash

:: checking this machine
    platform     Darwin/arm64
    disk free    612 GB at ~/.ommatidia/data
    memory       32 GB
    port         8787 available

    This downloads about 175 GB to ~/.ommatidia/data. Continue? [y/N] y

:: downloading the corpus (resumable; safe to interrupt)
    [214/214] shard-213.bin  812.4 MB  (48.2 MB/s)
:: connecting it to your agent
    OpenClaw       registered

Ommatidia is running.
    endpoint   http://127.0.0.1:8787/mcp/

Already have an MCP client? Skip the auto-registration and point it at http://127.0.0.1:8787/mcp/. It exposes three tools: search, get_passage and corpus_info.

Staying current

Updates that cannot become a tracking channel.

Design choice	Why
Static manifest, no query string	The mirror learns an IP fetched a public file, the same as any package index. No version, no install id.
Client never says what it has	It diffs locally. Sending "I have shards 1–213" would fingerprint the install.
Content-addressed shards	Requesting one reveals that shard and nothing else. Unchanged shards are never requested.
Signed and staged	Verified before use and swapped in atomically, so an interrupted update leaves the old corpus intact rather than a broken index.
Off is a real setting	`ommatidia update --schedule off`. An install that never updates keeps working forever.
Physical media supported	`--from /media/usb` applies the same signed deltas on a machine that has never seen a network.