Cherry Studio Workflows: Documents and Tools | Part 3

Practical AI with Cherry Studio and Omnirouter — Part 3 of 4.

The useful step beyond chat is not an unlimited agent swarm. It is a repeatable process with the right sources, a clear output and a human checkpoint. These Cherry Studio workflows start with one small document task and add tools only when they solve a specific problem.

Complete Part 2: provider setup first. A short chat response should work before you troubleshoot retrieval, tool calls or Agent protocol support.

Assistant, knowledge base and Agent are different things

A reusable assistant keeps role instructions and model settings together. A knowledge base organizes material that can be retrieved into a conversation. An Agent can carry out a sequence of actions using enabled tools. Do not treat these as interchangeable names for the same feature.

Cherry's Chat documentation describes assistants; its Agent documentation covers multi-step work. Exact controls depend on your installed version.

Start with an assistant when the task only needs reading and writing. Move to an Agent when the task genuinely needs actions, such as creating a new local deliverable or consulting an authorized service.

Workflow 1, then 2: build a small knowledge base and draft from it

Use a sanitized project FAQ, a public changelog and a short product overview. Three coherent documents are better for a first test than a folder full of duplicates and outdated notes.

1. Create a knowledge base and choose retrieval settings.
2. Import the files or supported sources.
3. Wait for processing to complete and check for failed items.
4. Test a question whose answer you can locate in one document.
5. Attach the knowledge base to the relevant conversation or assistant.

The current knowledge-base guide offers an embedding-free option, Do Not Use, with BM25 keyword retrieval. That is a useful starting point when it exists in your version. For semantic retrieval, configure a compatible embedding service or supported local embedding model. A working Omnirouter chat key does not establish embedding support.

Optional reranking is a separate component. Changing embedding models after indexing may require rebuilding the vectors; do not assume embeddings from different models are interchangeable. Consult retrieval configuration.

Now ask your assistant:

Using only the attached material, draft a support answer to this question. Identify the source section for every product claim. Separate confirmed facts from missing information. If the documents do not answer the question, say so instead of guessing.

Check the cited passages. A retrieval result can be incomplete or stale; a citation-looking string is not proof that the document supports the sentence. Ask a deliberately unanswerable question too. Correctly admitting a gap is part of a good result.

Workflow 3: prepare a new deliverable, not an uncontrolled edit

Where your version and provider support Agent work, start with a disposable working folder and copied sample material. Request a new Markdown brief rather than permission to reorganize your actual project.

Read only the supplied sample notes. Create a new release-brief.md containing shipped changes, limitations and open questions. Do not modify the source files. Do not run installation commands, send messages or publish anything. Ask before any action outside this scope.

Keep tool approvals enabled. Inspect proposed actions and review the resulting file against the notes. Once the small task is reliable, increase scope gradually.

Important compatibility boundary: Cherry's provider overview states that Agent access requires an Anthropic-compatible provider type. The compatible chat setup in Part 2 is not, by itself, an Agent setup. Confirm your Cherry release, gateway protocol and model support before trying this workflow through Omnirouter. You can still use the document-and-draft workflow in ordinary Chat while Agent compatibility is unresolved.

Add MCP only for a defined need

MCP connects supported tools and services to the client. According to Cherry's MCP guide, configuration, starting the server and binding it to an Agent are separate steps.

Use the server maintainer's documented configuration. Do not copy random commands into a tool configuration, and do not give a documentation lookup tool broad file or account access. A local MCP server runs software on your computer; a remote server may receive data you send to it.

For your first connection, choose a read-only service and one harmless task. Verify the exposed tools, authorize only the intended actions, then check the actual result. Connecting a service is not permission for an Agent to publish, delete, purchase or message people.

What Omnirouter contributes to these workflows

Omnirouter provides model inference for supported calls, not your local permissions, document index or third-party tool credentials. You choose a model that follows the task and works with the required protocol. For routine summarization and drafting, evaluate the GLM, DeepSeek, Qwen, Kimi and MiniMax options currently available in the catalog.

Cheap frontier access is useful for selective escalation, but upstream capacity and quality can change. Keep a tested alternative and make failures visible. Never hide a failed tool step behind a confident final answer.

Measure the outcome before adding autonomy

Record whether sources were followed, files were preserved, actions stayed in scope and the final result needed correction. A ten-step run that produces an unchecked answer is not better than a two-step run you can trust.

Next: Part 4: token costs and model selection. Return to Part 1: the workstation or Part 2: the connection.

Scope: Independent tutorial, not an official partnership. Feature labels and retrieval options vary by release. No universal Agent, embedding or tool compatibility is implied. Cover artwork illustrates a supervised workflow, not an application screenshot.