Back to Liquid Agent DOCUMENTATION

Your First Analysis

Start after installation. You do not need to write Python code to use the Web workbench.

For a complete real-data example, use the GSE174302 illustrated workflow: actual API answers, result-region questions, multi-image input and reviewed skill learning.

1. Open the Application

liquid-agent

The browser opens the local agent workbench directly. liquid-agent client is an equivalent entry point; liquid-agent web is also supported. Use liquid-agent cli for the terminal. The public homepage is for browsing before installation: its Try it link always opens the detailed installation Docs and does not check for a local workspace.

2. Configure GPT

Use the model selector and Manage keys to enter your OpenAI API key. Start with gpt-6-luna; gpt-6-sol and gpt-6-astra are explicit alternatives. The menu shows model names and key status. If authentication fails, inspect the error and key configuration rather than treating a failed response as a successful analysis. See GPT configuration.

3. Attach a Dataset

Choose New chat, select a local dataset folder, and wait for the scan to finish. Keep your data disk mounted while working. Scanning identifies available inputs; it does not mean an analysis has run.

For a complete example of the interface and review flow, follow the illustrated workspace walkthrough. Renaming a file does not make its assay compatible.

4. Ask, Review, Then Authorize

Begin with a focused request such as:

Inspect these data locally. Explain what they measure, which inputs are missing, and propose one useful QC analysis with a figure and a table. Do not run it yet.

Review the proposed inputs, prerequisites and plan. If it fits your question:

Run that QC analysis only. Include the figure, table and a short scientific interpretation in the report. Do not start a further analysis yet.

The precise plan depends on your files. A prerequisite question is not a failure: unsupported or ambiguous data must be clarified, not silently reinterpreted.

5. Inspect Results and Continue

Open Results and read the report, embedded images, tables and limitations. Then ask a question grounded in the output, for example:

Based on this QC, what should I examine next? Explain the evidence first and propose a new focused plan. Do not repeat the completed analysis.

New reports and plans retain earlier history. You decide whether to authorize the next analysis. See the Web guide for report history, changing a plan, editing a message and stopping an active task.

Before You Leave

Stop interrupts the active job, not the service. Closing the last local Web page or pressing Ctrl+C in the launching terminal can shut down the service. Keep a page open during a run. Deleting a conversation sends its owned records to Trash; restoring it is distinct from starting a fresh conversation.

The CLI guide provides the equivalent terminal workflow. CLI interface text remains English even when the website is Chinese.

If you use a local model

A cloud key is not required for a configured local profile. Use Manage local models, select its human-readable name, and check that its server is running. Follow local-model setup; installed weights and a successful tool probe do not establish scientific accuracy for every workflow.