Q&A Agent¶

WikiMind's Q&A agent answers questions against your compiled wiki, citing specific articles and suggesting follow-up questions.

Asking Questions¶

Simple question¶

curl -X POST http://localhost:7842/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What are the main arguments for and against microservices?"}'

The response includes:

{
  "query": {
    "id": "...",
    "question": "What are the main arguments...",
    "answer": "Based on your wiki articles...",
    "confidence": "high",
    "source_article_ids": ["Article Title 1", "Article Title 2"],
    "conversation_id": "..."
  },
  "conversation": {
    "id": "...",
    "title": "What are the main arguments..."
  }
}

Follow-up questions¶

Continue a conversation by passing the conversation_id:

curl -X POST http://localhost:7842/query \
  -H "Content-Type: application/json" \
  -d '{
    "question": "How does that compare to the monolith approach?",
    "conversation_id": "CONVERSATION_ID"
  }'

The agent uses prior turns as context to disambiguate references like "it" or "that approach". Up to 5 prior turns are included (configurable via WIKIMIND_QA__MAX_PRIOR_TURNS_IN_CONTEXT).

Streaming answers¶

For token-by-token streaming via Server-Sent Events:

curl -N http://localhost:7842/query/stream \
  -H "Content-Type: application/json" \
  -d '{"question": "Summarize my notes on distributed systems"}'

SSE events:

File-Back¶

When an answer has high or medium confidence, you can file it back to the wiki. This creates a new article from the conversation, making the wiki smarter over time.

Auto file-back¶

Set file_back: true in the request to automatically file high/medium confidence answers:

curl -X POST http://localhost:7842/query \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is the current state of quantum computing?",
    "file_back": true
  }'

Manual file-back¶

File an entire conversation back to the wiki:

curl -X POST http://localhost:7842/query/conversations/{conversation_id}/file-back

Or file selected turns from one or more conversations:

curl -X POST http://localhost:7842/query/conversations/file-back \
  -H "Content-Type: application/json" \
  -d '{
    "selections": [
      {"conversation_id": "...", "turn_indices": [0, 1, 2]}
    ]
  }'

Conversations¶

List conversations¶

curl http://localhost:7842/query/conversations

Returns conversations ordered by most recently updated first.

Get conversation detail¶

curl http://localhost:7842/query/conversations/{conversation_id}

Returns the conversation with all its turns (questions and answers).

Export as markdown¶

curl http://localhost:7842/query/conversations/{conversation_id}/export

Returns a standalone markdown document of the conversation.

Fork a conversation¶

Branch a conversation at a specific turn to explore a different line of reasoning:

curl -X POST http://localhost:7842/query/conversations/{conversation_id}/fork \
  -H "Content-Type: application/json" \
  -d '{
    "turn_index": 2,
    "question": "What if we approached it differently?"
  }'

This creates a new conversation that shares turns 0 through turn_index - 1 with the parent. The original branch is preserved immutably.

How Context Retrieval Works¶

When you ask a question, the Q&A agent:

1. Extracts key terms from your question
2. Scores all wiki articles by term overlap
3. Selects the top 5 most relevant articles
4. Truncates each to 3,000 characters to fit the context window
5. Includes prior conversation turns (up to 5) for multi-turn context
6. Sends everything to the LLM with a structured prompt

The LLM responds with JSON containing the answer, confidence level, cited sources, related articles, and follow-up questions.