CASPER (Cambridge Astronomical Survey Personal Expert Resource) is an assistant platform developed by CamCEAD for astronomical survey teams. It answers technical questions from curated project documents, shows where each answer comes from, and can carry out approved tasks on behalf of the people who use it.

Current status. CASPER already serves several projects. CamCEAD's own projects, CamCEAD and CAMCEADHUB, are part of it, and the science projects WEAVE and GAIA are on board in our pilot. It is used by CamCEAD and project team members, is in beta and is not open to the public.

Why CASPER

Large survey projects run on long specifications, procedures that change, data that passes through many systems, and knowledge that often sits with a few specialists. New team members spend weeks finding out which document is current, and experts are asked again and again for answers that are already written down somewhere.

CASPER is our working answer. It reads the material a project has chosen to share, finds the passages that bear on a question, and writes an answer that points back to them. It can also use the tools a team already has, with the same permissions as the person asking. It is at once a service that people use and a test bed for our own research: how far can an assistant help in a survey setting while every answer stays traceable to its source and every action stays under human control?

What CASPER offers

How CASPER is organised

CASPER is a platform of several layers. A request passes through each of them, and the rules of the first layer apply to everything below it.

• Access and policy. Every request carries who is asking and for which project. Sign-in uses the university account or a local account, roles and project membership decide what can be read and which tools can be used, and each answer and action is recorded.
• Orchestration. A router decides what kind of request it is (a knowledge question, a data question, an operations check, code or a helpdesk request) and hands it to the right specialist agent. Skills tell an agent how a team likes a task done.
• Knowledge, tools and models. Agents use three services: multi-layer retrieval for evidence, a family of tool servers for data and systems, and language models to read and write.
• Sources and systems. Documents, papers and code repositories, survey archives and pipelines, and the research computing provided through IRIS on which CASPER runs.

Knowledge: from a question to a cited answer

Project teams choose what CASPER may read: specifications and procedures, papers, software repositories, web pages and, where a project agrees, recordings turned into text. New material is checked for credentials, labelled with the project and the level of access it needs, and split into passages before it is added.

When a question arrives, CASPER first works out what is being asked, including any names, dates and identifiers in it. It then searches in five complementary ways at once, because no single method works for every question. The results are merged and re-ranked, and the answer is written only from the best evidence. Each citation is checked against the passage it points to.

• Meaning finds passages that say the same thing in different words.
• Keywords finds exact terms, instrument settings and identifiers that meaning-based search can miss.
• Whole documents uses summaries to find the right file before the right paragraph.
• The knowledge graph follows links between instruments, processing steps, teams and documents, and answers questions that span several sources.
• Time and people narrows a search to a period, an author or a channel.

The knowledge base also examines itself. Every night it asks itself test questions, and a set of reference questions with known answers is used to measure every change before it is adopted.

Tools: a family of MCP servers

The Model Context Protocol (MCP) is an open standard that lets an AI assistant call tools in a controlled way. CamCEAD designs and builds its own family of MCP servers, with different tools for different projects: searching archives, reading file headers, checking the state of a data flow, looking up project knowledge, and handling support tickets.

People can reach all of this from the CASPER website or from their own AI desktop client, which needs one address and one sign-in. The tools a person sees depend on their role and projects. Almost all of them read, inspect or prepare a plan. Anything that would change a system needs an administrator and an explicit confirmation, and CASPER never starts or restarts processing on its own.

With more than three hundred tools, showing every one of them to a model is slow and expensive. CASPER is therefore gaining a tool finder: the model describes what it needs and is offered the right tools, and only tools that the person is allowed to use. A tool run through the finder passes exactly the same checks as a direct call. Results of tools that only read can be reused for a short time without ever mixing different people's data, and very large results are kept for the person who asked and read back in parts. Every week the wording of each tool is scored, and clearer wording is proposed to a person who decides whether to adopt it. These parts are being introduced step by step and are switched on only after they have been measured.

Models: commercial and open, chosen by task

CASPER does not depend on a single model. A router chooses a model lane for each request by the task and by how sensitive the project is. Commercial models handle the hardest work, such as long documents, careful reasoning and written reports, and only for projects that allow it. Open models, hosted and managed by CamCEAD on research computing, do the routine work: analysing questions, indexing documents, reviewing answers and turning recordings into text. A project can be marked local-only, and then every question about it is answered entirely on hardware we run ourselves.

The open models we host today are from the Mistral family and from OpenAI's open-weight family. They replaced our earlier open models in October 2026. Each model is measured on our own reference questions before it replaces another, and the platform is model-agnostic, so a change is made one lane at a time without any change for the people who use CASPER. Models are not trained on user conversations.

Projects on CASPER

CASPER is built for many projects at once. Each project has its own lane on the shared platform: its own knowledge, tools and assistants, its own people and roles, and its own privacy setting. Two kinds of project use it today: CamCEAD's own platforms, and science projects in the pilot.

Built to be trusted

• Project separation. The knowledge of different projects is kept apart, and the separation is enforced in the search, the tool servers, the agents and the pages. An administrator can manage only their own projects.
• Grounded answers. Answers carry their sources, citations are checked, and CASPER says when nothing supports an answer.
• Human control. Changes to a system need an administrator and an explicit confirmation. Shared knowledge from a personal collection needs the owner's consent first.
• People in the loop. 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. This is being introduced and is not yet in general use.
• Records. Answers and actions are recorded so that they can be reviewed, and changes to a project's knowledge are logged.
• Strict public assistants. Assistants on public websites answer only from pages that anyone can read, refuse attempts to redirect them, and limit what each visitor can ask.
• Privacy. Models are not trained on user conversations. When a commercial model is used it receives only the text needed for that answer. Projects that need extra care are answered entirely on hardware we run.
• Testing. New models, tools and settings are measured against reference questions before they are adopted.

Next steps

The pilot with WEAVE and GAIA shows what works and what does not. We keep measuring our hosted models against our own reference questions, we are widening the tools available to each project, and preparing a peer-reviewed paper that will describe the system in full. We are also preparing a stand-alone version of our astronomy tools, with a benchmark for how well AI models use such tools in astronomy, with the aim of sharing both openly. Once the pilot has shown what works, we plan to extend CASPER to further surveys and programmes with similar needs.

Limits

CASPER can be wrong, so its answers carry sources that you should check. Anything that would change a system needs an administrator and an explicit confirmation, and CASPER does not start or restart processing on its own.

Access

CASPER is available to authorised users. To ask for access, please email casper@ast.cam.ac.uk.

Sign in to CASPER