Knowledge Q&A

Knowledge Q&A combines a large language model with knowledge retrieval to fetch relevant content and generate a natural-language answer. It supports multi-Knowledge-Base binding, multi-turn dialogue, streaming output, and citation display.

Capabilities

Capability	Description
Intelligent Q&A	The LLM combines retrieval results with its own reasoning to produce cited answers.
Multi-Knowledge-Base binding	Bind up to 15 Knowledge Bases with per-source priorities and retrieval parameters.
Two retrieval modes	Fast Mode (single-pass retrieval) and Multi-turn Agentic Mode (Agentic multi-step planning).
In-chat file parsing	Upload a file mid-conversation; the system parses it on the fly without importing it into a Knowledge Base.
Generation controls	Refusal policy, prompt-leak protection, and citation rendering.

Create a Q&A service

Open Knowledge Q&A

Create a service

Click New Q&A Service in the top right and fill in a name and description.

Configure model and prompt

Pick a generation model (such as qwen3-plus or DeepSeek-R1).
Optionally enter a prompt (up to 500 characters) to steer the model's tone.

Pick a retrieval mode

Choose Fast Mode or Multi-turn Agentic Mode.

Bind Knowledge Bases

Click Add to bind Knowledge Bases (up to 15) and configure per-source retrieval parameters.

Configure generation controls

Enable the refusal policy, prompt-leak protection, and other controls as needed.

Publish

Click Publish in the top right.

Retrieval modes

Mode	How it works	Best for
Fast Mode	Single-pass retrieval with the lowest latency.	Simple factual questions.
Multi-turn Agentic Mode	The model plans the search strategy, rewrites the query multiple times, and merges results.	Complex questions, multi-step reasoning, information spread across documents.

Generation controls

Parameter	Description
Prompt	Steers the model's tone — for example, "Answer only based on the Knowledge Base."
Refusal policy	If nothing relevant is retrieved, the model declines instead of fabricating an answer.
Prompt-leak protection	Prevents the model from echoing the system prompt or internal instructions.
Citations	Adds citation markers in the answer that link back to the source Chunk.
Multimodal reply	Answers can include images, tables, and other rich media.

API calls

Knowledge Q&A is invoked through the chat endpoint and returns an SSE stream:

curl -X POST https://{workspaceId}.cn-beijing.maas.aliyuncs.com/api/v2/apps/knowledge/chat \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "input": {
      "messages": [
        {"role": "user", "content": "How do I choose a chunking strategy?"}
      ]
    },
    "parameters": {
      "agent_options": {
        "agent_id": "${YOUR_AGENT_ID}"
      }
    },
    "stream": true
  }'

SSE event types

The Q&A response streams these events in order:

Event	Description
`plan_start` → `planning` → `plan_end`	Planning — the model plans the retrieval and answer strategy.
`tool_calling` → `tool_return`	Tool calling — `tool_calling` emits the retrieval request; `tool_return` returns matched chunks.
`generation_start` → `generating` → `generation_end`	Generation — streams the final answer; `finish_reason` is `stop` at `generation_end`.

In Multi-turn Agentic Mode, the plan → tool_calling → tool_return segment can repeat multiple times. See the chat endpoint for full event details.

Multi-turn dialogue

The platform doesn't persist conversation state — pass the full messages history on every request. Recommendations:

Cap history length (around the last 10 turns) to stay within the model's context window.
Retrieve based on the most recent user message only — older messages add noise.

After publishing the Q&A service, check Service Channels for integration options via API, MCP, CLI, and more.

​Capabilities

​Create a Q&A service

​Retrieval modes

​Generation controls

​API calls

​SSE event types

​Multi-turn dialogue

Capabilities

Create a Q&A service

Retrieval modes

Generation controls

API calls

SSE event types

Multi-turn dialogue