DB-GPT

Use this repo skill when a task names DB-GPT, dbgpt, dbgpt-app, dbgpt-client, dbgpt-ext, dbgpt-serve, or dbgpt-sandbox, or asks for an agentic data assistant that combines models, SQL/code, files, knowledge bases, RAG, skills, AWEL workflows, APIs, or sandboxed execution.

Choose the route

• Install, profiles, TOML configuration, dbgpt command routing, web start/stop, knowledge/repo/app/trace/migration CLI → setup-and-cli.
• LLM/embedding/reranker providers, model TOML, controller/worker/API-server, local models, CUDA, vLLM, llama.cpp, Ollama, or serving → models-and-serving.
• Agents, tools, skills, middleware, teams, memory/context, AWEL/DAG/flow, or HTTP-triggered workflow construction → agents-and-awel.
• Local documents, datasource schemas, chunking, embeddings, retrieval, knowledge indexing, vector/full-text/graph stores, or RAG → data-and-rag.
• Python client, HTTP CRUD/API routes, files/apps/flows, service integration, or sandbox execution → apis-client-and-sandbox.

When a request spans routes, establish the runtime boundary in this order: configuration/provider → data or workflow construction → API/service execution. Do not claim that a parsed config, imported package, registry entry, or health endpoint proves a provider, model, vector store, database, or sandbox is usable.

Baseline installation and inspection

For the recorded 0.8.1 baseline, use a clean Python 3.10+ environment and install the public package without relying on a source checkout:

uv pip install 'dbgpt-app==0.8.1'
dbgpt --version
python -c "import dbgpt, dbgpt_app, dbgpt_client, dbgpt_ext; print('imports ok')"

Use pip if uv is unavailable. Add only the extras needed by the route: provider, RAG/parser, vector-store, database, local-model, or sandbox variants. Read installation before adding optional backends, and configuration before writing a profile or TOML file. The bundled package smoke script is a read-only import/version check; it never starts a server or contacts a provider.

Cross-cutting operating rules

1. Keep credentials in environment variables or the user's secret manager. Do not print resolved TOML, API keys, passwords, bearer tokens, or provider responses containing secrets.
2. Separate local/import/config validation from live provider, model, database, vector, graph, API, and sandbox checks. Record the exact boundary and skipped prerequisites.
3. Prefer tiny deterministic fixtures and disposable paths. Do not run a network installer, download a model, mutate a database, launch a daemon, or execute untrusted code as an implicit smoke test.
4. Use the installed package's live Click help and signatures when a flag or object is version-sensitive. In particular, model command discovery may query a controller even while rendering help; an unreachable controller is not a successful deployment.
5. Default DB-GPT's web API is commonly http://localhost:5670; confirm the actual profile, API version, route prefix, and authentication policy before making client calls. Close async clients and clean temporary uploads/sessions.
6. Treat local sandbox execution as a weaker, explicitly opted-in boundary, not container isolation. Container runtime/image availability and network policy must be verified separately.

Read cross-cutting troubleshooting for installation/import, optional dependency, configuration, service, backend, and safety failures. Source/version context is recorded in repo provenance; it is not a runtime dependency.