---
title: "AI research assistant: ask your files and the literature"
description: "Biocompile's AI research assistant searches PubMed and bioRxiv, reads your CSV, FASTA and PDF files, runs Python and R code and proposes project changes."
url: https://www.biocompile.com/docs/ai-research-assistant/
section: "Projects and assistant"
last_updated: 2026-10-23
app_url: https://app.biocompile.com/conversations
---

# AI research assistant: ask your files and the literature

> Biocompile's AI research assistant searches PubMed and bioRxiv, reads your CSV, FASTA and PDF files, runs Python and R code and proposes project changes.

Source: https://www.biocompile.com/docs/ai-research-assistant/. Index of all Biocompile documentation: https://www.biocompile.com/docs/index.md

Biocompile's AI research assistant is an AI co-scientist that answers questions about your
files and the literature, runs analyses as Python and R code in a private sandbox, and
proposes changes to your projects as contributions. The Biocompile homepage calls it the AI
co-scientist. Everything it produces stays in the conversation until it is contributed to
a project.

## What the assistant can do

- Search the literature in PubMed, arXiv, bioRxiv, OpenAlex and Semantic Scholar, search patents by class, applicant and country, query public biological databases, and cite every source it uses.
- Read files from your Knowledge Atlas when you mention them.
- Write and run analysis code in an isolated sandbox with Python, R and Bioconductor, and show you the code it ran.
- Create files such as tables, figures and notes that you can reuse later.
- Pick up earlier work in the same conversation and build on it.

## Work in a conversation

A conversation with the assistant opens in the assistant pane. If the pane is hidden,
choose **Show assistant** in the top bar, then choose **New conversation**. **Conversations**
in the sidebar lists every conversation with its project and last activity.

Ask a focused question and name the files that matter. In the message box, type `@` to
reference a file, or drag or paste a file to attach it. The assistant works in visible
**Work steps**: open a step to see its **Request** and **Response**, including the code it
ran.

The send button stays disabled until the message has text or an attachment. When a
selection you made answers the assistant, a note says it is ready and asks you to press
**Send**.

> Fit a logistic growth model to each well in plate_reader_OD600_2026-03.csv and flag
> wells whose curves suggest contamination.

Each conversation has a **Conversation workbench** that collects what the assistant
produced: proposed project changes, the files and sandboxes it used, the research material
it found, and its activity.

## If a sandbox does not start

When the sandbox cannot start, a **Sandbox could not start** notice explains why and what to do:

| Cause | What the notice tells you |
|---|---|
| Quota | Your sandbox resource quota is full. Stop unused sandboxes or ask your administrator to review your quota. |
| Capacity | Sandbox resources are busy. Wait a moment, then try again. |
| Timeout | The sandbox did not start in time. Try again, and contact support if it continues. |
| Recovery | The sandbox could not be recovered safely. Contact support before you try to restore it again. |

Except for a recovery problem, the notice offers **Try starting again**.

## Link a conversation to a project

A conversation works with one project. Under the message box, open **No project** and
select a project under **Project context**. The first selection binds the conversation to
that project for good, and so does the first change the assistant proposes to one. Start a
new conversation to work with a different project.

Selecting a project gives the assistant its context and fixes where proposed work will go.
It does not add anything to the project by itself. A project whose agent change policy is
**Read only** can still provide context, but cannot receive contributions.

## Choose the model

The assistant's model comes from the agent profile you choose beside the message box
(**Choose co-scientist**). Each profile has a default model family, set under
**Agents → Agent Profiles**. Within that family, **Choose intelligence** sets the model
size and thinking for the conversation. The agent profile is fixed once a conversation
has started, so start a new conversation to use another profile.

Your organization's security policy decides which models you may use. A model it does
not allow is marked "Blocked by organization policy". See
[When a model is blocked](https://www.biocompile.com/docs/organization-security.md#when-a-model-is-blocked).

## Contribute conversation work

A contribution is the reviewed set of changes through which work from a Biocompile
conversation reaches a project. Changes the assistant proposes are provisional: they stay
in the conversation's **Contribution draft** until they are contributed and accepted.
Hovering over **Contribute work** shows **How contributions work**, a summary of the flow.

1. Choose **Contribute work** under the message box. When the assistant asks
   "Contribute this work?", choose **Contribute work**, or **Not now** to decline.
2. If the conversation is not bound to a project yet, choose one under **Choose a project**,
   then choose **Prepare contribution**. Result: the conversation shows "Preparing
   contribution" while Biocompile collects the durable findings into proposed changes,
   and the message input is paused.
3. Review the contribution: its summary, every proposed change as a before-and-after
   comparison, and the evidence behind each one. You can edit proposed values; your edits
   are saved as a new revision, and the version you reviewed first stays in the history.
4. Choose **Accept contribution**. Result: Biocompile shows "Contribution approved and
   applied", and the changes appear in the project.

The same review is available on the project's **Contributions** tab. Only accepting a
contribution changes the project; proposing and preparing never do.

### Request a revision

Choose **Revise**, describe what should change under **Revision instructions**, and choose
**Request revision**. Choose **Withdraw** to drop your own contribution, or **Decline** for
someone else's.

### When a contribution cannot be applied

A conflicted contribution applies nothing, and the review names the reason:

- **A project rule prevented this contribution.** The change would break a rule of the
  project, such as two records with the same name. Revise the proposed changes and submit
  them again.
- **The project changed before approval.** A value the contribution relied on changed
  after the contribution was prepared. Choose **Revise conflicts** and revise the proposed
  changes using the conflict details.

### Auto Mode

**Auto Mode** sits beside the message box. It lets the assistant answer the approval
checkpoints you select for it under **What Auto Mode handles for you** in
**Settings → Assistant**. With **Contributions** selected there, it prepares and applies
contributions without asking. With Auto Mode off (Contribute Mode), every contribution
waits for your review.

An organization admin can pin a category for everyone under **Organization → Policies**,
for example **Always review**. Demo and regulated (GxP) organizations cannot use Auto Mode.

## Common questions

### Can the assistant change my project without my approval?

Only if you allow it. In Contribute Mode, nothing the assistant proposes reaches a project
until someone chooses **Accept contribution**. In Auto Mode with **Contributions** selected,
the assistant applies contributions itself. To review each one, turn Auto Mode off in the
conversation, or ask an admin to set contributions to **Always review**. A project whose
agent change policy is **Read only** never receives contributions.

### Which AI models can I use in Biocompile?

Biocompile offers Claude and GPT models and open models such as Qwen, DeepSeek and GLM, in
model families with smaller and larger members. Some families process data in the EU. The
agent profile you choose sets the family, and your organization's security policy decides
which models are available.

### Does the assistant show the code it ran?

Yes. Each step of the assistant's work shows its request and response, so you can read the
code behind every result.

## Related

- [Why AI models decline biology questions, and what to do](https://www.biocompile.com/docs/model-declines.md)
- [Agent profiles: persona, skills and memory for the assistant](https://www.biocompile.com/docs/agent-profiles.md)
- [Research reports: save cited findings and build on them](https://www.biocompile.com/docs/research-reports.md)
- [Research projects: hypotheses, evidence and decisions](https://www.biocompile.com/docs/projects.md)
- [Knowledge Atlas: upload, preview and edit lab files](https://www.biocompile.com/docs/knowledge-atlas.md)

## Get access to Biocompile

Biocompile is used in the browser at https://app.biocompile.com. It is free to start and no payment method is required: sign up at https://app.biocompile.com/signup?mc_cta=docs-md&mc_path=/docs/ai-research-assistant.md. If your team already uses Biocompile, ask an organization admin to invite you. Research teams that want hands-on setup can apply for the founder-led pilot at https://www.biocompile.com/pilot/. Questions: https://www.biocompile.com/contact/.