---

name: "deep-analyze" description: "Router for DeepAnalyze agentic data-science workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

DeepAnalyze

DeepAnalyze is an agentic data-science repo built around a local reasoning loop, an OpenAI-compatible API server, browser and notebook frontends, model-serving helpers, and official training/evaluation recipes.

Use this router when the request names DeepAnalyze, DeepAnalyze-8B, the OpenAI-compatible API, the browser demo, the CLI, Jupyter, vLLM serving, quantization, tokenizer customization, SFT/RL recipes, or the bundled benchmark playgrounds.

If the checkout may be stale, consult references/repo-provenance.md before trusting any route or claim.

Start here

1. If you are setting up or checking an environment, read references/install-and-environment.md.
2. If something fails, read references/troubleshooting.md.
3. For a read-only health check, run scripts/check_deepanalyze_environment.py against a DeepAnalyze checkout.
4. Then route to the sub-skill that owns the task surface.

Route map

Task family	Read first	Why
Programmatic agent loop, workspace execution, file-aware client flows	sub-skills/api-and-clients/SKILL.md	Covers DeepAnalyzeVLLM, OpenAI-style clients, file/thread semantics, and API smoke scripts.
CLI, browser WebUI v2, or Jupyter frontend	sub-skills/interactive-frontends/SKILL.md	Covers the terminal client, workspace UI, Docker execution mode, and Jupyter MCP flow.
Model download, vLLM launch, quantization, or tokenizer tag extension	sub-skills/model-serving/SKILL.md	Covers the memory table, dry-run vLLM commands, Docker deployment, and model customization.
SFT, RL, benchmark runs, or case-study contribution	sub-skills/training-and-evaluation/SKILL.md	Covers the official DeepAnalyze training recipes and benchmark playgrounds.

What this router should answer

• How to call DeepAnalyze from Python or through an OpenAI-compatible client.
• How to upload files, keep a thread workspace, and interpret generated files.
• How to launch or preflight the browser and notebook experiences.
• How to choose a DeepAnalyze vLLM command from GPU memory and context requirements.
• How to plan tag-extension, quantization, SFT, RL, or benchmark commands without running placeholder-heavy jobs blindly.

Minimal read-only check

If you need a quick confidence check without starting services, use the bundled checker:

python scripts/check_deepanalyze_environment.py --repo-root <DeepAnalyze checkout>

That checker verifies the core Python imports, selected source-file compilation, DeepAnalyzeVLLM, API TestClient behavior, and WebUI v2 TestClient behavior.

Route away from this root when

• You already know the exact sub-skill and only need its detailed reference.
• The task is only about the older legacy browser demo, unless the user explicitly asks for it.
• The task is about a different package's agent framework, model server, or benchmark harness.

Preferred mental model

• api-and-clients = "How do I use DeepAnalyze programmatically or test the API?"
• interactive-frontends = "How do I use the CLI, browser demo, or notebook UI?"
• model-serving = "How do I size, serve, or customize DeepAnalyze-8B?"
• training-and-evaluation = "How do I dry-run training, RL, or benchmark jobs?"

When a request spans multiple areas, pick the owner of the primary workflow first and then follow its cross-links.