What CASPER does

Answers with sources

CASPER answers technical questions from the documents a team has chosen to share, such as specifications, procedures, papers, software repositories and web pages. Where a team agrees, recordings turned into text can be included. New material is checked for credentials, labelled with its project and the level of access it needs, and split into passages before it is added.

When a question arrives, CASPER works out what is being asked, searches in several complementary ways at once (by meaning, by exact keywords, by whole documents, through a knowledge graph, and by time and author), merges and re-ranks the results, and writes the answer only from the best evidence. Each citation is checked against the passage it points to. When nothing supports an answer, CASPER says so instead of guessing.

Expert tools

Registered members can access a project based set of expert tools. They read, inspect and prepare work on data and systems, for example looking up information, checking a state or preparing a plan. They run with the permissions of the person asking, and the permissions are checked again when a tool runs, not only when it is listed.

Anything that would change something needs an administrator and an explicit confirmation. CASPER never starts or restarts work on its own. Tools are built by CamCEAD on the open Model Context Protocol, so they can also be reached from an AI client a person already uses.

HPC services and tools

CASPER also looks after the research computing that analysis runs on. Registered members and administrators can see the state of a machine: the job queue, the nodes, usage and health, and why a job is waiting, each person seeing only what they may. The tools help with performance and optimisation, showing how jobs use processors, memory and GPUs, where capacity sits idle and what to change.

Anything that changes a machine is for administrators, starts with a dry run, needs a typed confirmation and is recorded, and protected systems are refused.

AI governance and compliance

CASPER gives an institution control over how AI is used. Code and projects can be checked against the security policy before they go in, documents and requests are screened for credentials and for text that tries to instruct the assistant, and each project has its own rules for which models may be used and whether anything may leave CamCEAD.

Every answer and action is recorded for review, usage and cost are reported to administrators, retention periods are set, and the written acceptable use terms say what is and is not allowed.

Specialist helpers

A router decides what kind of request it has received and hands it to a focused assistant, for example for literature search, code, data, operations or support requests. Each assistant is limited to its own project and task, and has its own list of tools and its own budget.

Assistants inside websites

Short, strict assistants can live inside a website. They answer only from pages that anyone can read, refuse attempts to redirect them, and limit what each visitor can ask. They give a polite pointer to CASPER for anything else.

Helpdesk

First-line help for common problems, and a drafted support ticket that the person reads and approves before it is sent.

Skills

Teams can teach CASPER how they like a task done, once, and have it done that way every time. Suggested improvements are reviewed by a person before they are adopted.

Command line

A small program, casper, lets a registered member use CASPER from a terminal or a script, with their own account. It sees exactly what that account sees and adds no rights. It also has a check that says what is wrong with a set-up and how to fix it.

Tool finder

With hundreds of tools, a person should not have to know their names. The tool finder lets an assistant describe what it needs and be offered the right tools, only those the person is allowed to use. A tool run through the finder passes exactly the same checks as a direct call. Large results are kept for the person who asked and read back in parts, and results of tools that only read can be reused for a short time without ever mixing different people's data.

Project memory

CASPER can keep, for each project, four different things apart: verified facts, decisions with their reasons and history, temporary context that expires by itself, and a person's own preferences. Each entry has its source. A past conversation is never treated as authoritative, and nothing is overwritten silently.

Ask an expert

When CASPER is unsure and no authoritative source exists, it can offer to pass the question to a named expert of the project, but only with the asker's agreement. The reply is shown as coming from that person, and not as something CASPER has verified.

Open astronomy tools

A read-only set of astronomy tools for AI agents and scripts, with a benchmark for how well language models use them, is available to every registered member now, and is being released as an open package. See the tools and the benchmark.