[Template]: ragflow-knowledge-qa-agent

[Everfighting]

Template name

ragflow-knowledge-qa-agent

What does this template do?

Traditional solutions integrating RAGFlow Skills have limited flexibility. This template encapsulates RAGFlow dataset access and document chunking as custom skills. It fully leverages the agent's task planning, dynamic workflow generation and autonomous tool calling capabilities, combined with native memory and logical reasoning, to build a highly flexible knowledge base Q&A application free from rigid predefined workflows.

Who is this for?

Developers (building/extending agents)

Agent framework / harness

Not sure — recommend one

Capabilities to support

• 🔍 Retrieval / RAG (OceanBase · seekdb vector + fulltext)
• 🔐 Auth (per-user / resource-level authorization)
• 🧠 Memory & context (ContextSeek / persistent context)

🛠️ Skills — details

• Reused: Leverage native capabilities of RAGFlow, including document parsing, document chunking and dataset retrieval
• New: Session context memory, text summarization, answer generation, logical reasoning, dynamic task planning, autonomous tool calling, general LangSmith tracing, user identity verification, knowledge base scope filtering by permissions
• Approval rules: Manual review and editing are required before bulk dataset operations and large-scale document chunking, Access across knowledge base permission scopes will be blocked and alerted

🔌 MCP — details

No response

🧑‍⚖️ Human-in-the-loop — details

• Pause for user approval before invoking RAGFlow for bulk document chunking and dataset operations.
• Support interrupting ongoing tasks to manually adjust retrieval scope and modify query instructions.
• Allow editing of agent-generated answers and execution strategies, and resume tasks after modification.
• Mandate manual review for Q&A scenarios involving confidential knowledge bases and unauthorized access attempts to knowledge bases.

📡 Event streaming — details

No response

Models / providers

No response

How will it run?

None

External services & data

No response

What does success look like?

No response

Priority

None

Are you able to help build it?

Yes — I'll open the PR

Anything else?

No response

[webup]

Thanks for the detailed proposal, @Everfighting — this is a great fit and we'd be happy to take it as a contributed template. A few things to align on before you open the PR.

Framework → DeepAgents. Your pitch is explicitly "task planning + dynamic workflow generation + autonomous tool calling, free from rigid predefined workflows" — that's the DeepAgents value proposition, so let's base this on the deepagents/ family (closest existing template: deepagents/research, which already ships tool-calling + a streamed tool/sub-agent UI). Proposed spec: deepagents/ragflow-knowledge-qa.

OceanBase seekdb as the recommended backend. RAGFlow supports seekdb as its document engine (DOC_ENGINE=seekdb, see ragflow#12691 and this OceanBase walkthrough), so the "🔍 Retrieval / RAG (OceanBase · seekdb)" capability is fully achievable. For an AgentSeek template we'd strongly prefer seekdb as the recommended, first-class default rather than one option among ES/Infinity — it keeps the whole stack on a single AI-native database (vector + fulltext + hybrid in one engine), aligns with AgentSeek's database-native thesis, and lets the agent's RAG chunks, runtime data, and ContextSeek memory all converge on OceanBase.

A few scoping questions:

1. RAGFlow deployment model — does the template assume a user-provided running RAGFlow instance (configured via .env: base URL + API key, plus seekdb conn), or do you want to bundle a docker-compose that brings up RAGFlow + seekdb? We'd lean toward bring-your-own RAGFlow endpoint to keep the starter light, with the seekdb compose documented as the recommended setup path — but open to your view.
2. Tool surface — which RAGFlow operations do you plan to expose as agent tools/skills? Mapping to ragflow-sdk (list_datasets / retrieve / upload_documents / parse_documents) is the natural path. Confirm the set.
3. HITL scope — you list approval gates before bulk chunking/dataset ops and interrupt/edit/resume of retrieval scope. Which of these are v1 vs follow-up? Approval-before-bulk-ops is a clean v1; full interrupt-edit-resume is a larger lift.
4. Models/providers — none specified. What's the intended default LLM + embedding/rerank model for the agent side (RAGFlow's own model config is separate)?

On build: it'd live at templates/deepagents/ragflow-knowledge-qa/ (cookiecutter), registered in templates/index.json + docs/reference/templates.md. Please follow CONTRIBUTING.md (tests + make check, docs update). Since you marked "I'll open the PR" — great, looking forward to it.

[Everfighting]

Thanks for the clarification,Here are my confirmations:

1. The template assumes users use their existing deployed RAGFlow instance. Existing RAGFlow deployments will be guided to use SeekDB as the default backend.
2. Agreed. We'll expose all four listed operations.
3. Agreed. Bulk approval will be in V1, and interrupt-edit-resume will come later.
4. No default agent models. All follow user's RAGFlow config.