Reducing Wall-Clock Latency and Token Spend for Large-Scale Data Access
Public scientific archives are large and irregular, and an AI agent spends most of its round trips to the model working out where things are and how to read them. That work is the same for every user, so it can be done once and handed to the agent as a schema card. On five archives from 255 MB to about 515 TB, the schema card reduced round trips per question from 6 to 2 and tokens by 30 percent, with the same accuracy.
Public scientific archives are large and irregular. A single release can hold hundreds of thousands of files in a layout specific to one lab, in formats that need a particular library, with conventions documented in a paper or nowhere at all. Before anyone can answer a question with the data, they have to work out where the relevant file is, how to open it, and what the values mean.
More of that work is now done by AI agents, in notebooks and in automated pipelines. An agent works out an unfamiliar archive the way a person would: it lists directories, opens sample files, installs libraries, and retries when something fails. Each of those steps is a round trip to the model. Round trips are what make an agent slow, and the tokens they consume are what make it expensive. On a large archive, a single listing can take minutes.
What a schema card contains
The agent decodes it and renders the schema card into its system prompt before the first question. Outside a notebook, the catalog API and the MCP server return the same JSON.
Setup
Four questions per archive, 20 in total, each with a numeric or list answer that Scigantic computed beforehand. Some examples:
Each question ran twice in each of five setups with the production agent code and model (Claude Opus 5). A second model graded the answers against the precomputed ones.
Schema card alone
37 of 40 answers were correct without the schema card and 36 of 39 with it. In 37 of 39 paired questions, the run with the schema card made fewer round trips. No answer repeated the schema card in place of reading the data.
Container image and FUSE mount
All five setups:
On the general Python image the agent ran pip install in 26 of 40 runs. GFS is stored in Icechunk, which needs Python 3.12; the general image has 3.11. There the agent used 305 round trips and 42 minutes across eight questions and never opened the data.
Without the mount, the schema card was passed in the user message, where it is not served from the prompt cache, so cost per question stayed at $0.11 despite fewer tokens. In production the schema card is in the cached system prompt.
Pipelines
A batch job or workflow step that calls an agent per item can fetch the schema card from the catalog API or MCP server. Round trips and tokens are then paid per item. With the schema card and no mount, the 90th percentile wall clock was 78 s and no run hit the step limit. With neither, it was 404 s and two runs hit the limit without an answer.
Scigantic intends to make schema cards more robust and to learn how best to navigate very large object stores. That means covering more archives, keeping each schema card current as the data underneath it changes, and testing with more questions and more agents running at the same time.