Skip to content

qa.md

Latest commit

 

History

History
131 lines (94 loc) · 2.91 KB

qa.md

File metadata and controls

131 lines (94 loc) · 2.91 KB
Raw

Q&A

Duct can answer questions about your documents using a configurable LLM provider. The process is: search for relevant context, send it to the LLM with your question, return the answer with source citations.

Quick Start

# With Ollama (default — no API keys)
duct ask "What are the termination clauses?"

# With OpenAI
duct ask "Summarize the indemnification" --llm openai

# With Gemini
duct ask "List all parties" --llm gemini

Providers

Ollama (Default)

Runs locally — no API keys, no data leaving your machine.

duct ask "What is the governing law?" --llm ollama
duct ask "..." --llm ollama --model llama3.2 --base-url http://localhost:11434
Option Default
Model llama3.2
Base URL http://localhost:11434

Requires Ollama to be running locally: ollama serve

OpenAI

export OPENAI_API_KEY=sk-...
duct ask "Summarize the agreement" --llm openai
duct ask "..." --llm openai --model gpt-4o
Option Default
Model gpt-4o
Base URL https://api.openai.com/v1

Gemini

export GEMINI_API_KEY=...
duct ask "What are the obligations?" --llm gemini
duct ask "..." --llm gemini --model gemini-2.0-flash
Option Default
Model gemini-2.0-flash
Base URL https://generativelanguage.googleapis.com/v1beta

Configuration

Set the LLM provider in the web UI Settings panel, or via the CLI:

# Just ask once with a specific provider
duct ask "Question" --llm openai

# Start the server with a default LLM
duct serve --llm ollama

Agentic Retrieval

For complex questions that span multiple documents or topics, agentic retrieval decomposes your question into sub-queries, searches each independently, and synthesizes the final answer.

duct ask "Compare the indemnification clauses in all contracts" --multi

This is useful for:

  • "What are the differences between the two NDAs?"
  • "Which contracts have a termination for convenience clause?"
  • "Summarize all payment terms across vendors"

Context-Only Mode

Skip the LLM call and see what context would be retrieved:

duct ask "What are the termination clauses?" --no-answer
duct ask "Question" --no-answer --json    # JSON output

Output

By default, the answer is printed with source citations:

  Searching for: "What are the termination clauses?"

  Answer (1523ms):

  The contract includes a 30-day notice period for termination without cause...

  Sources:
    [9.2] contract.pdf > Termination
    [7.8] appendix.pdf > Term

Use --json for machine-readable output:

duct ask "Question" --json
{
  "answer": "The contract includes a 30-day notice period...",
  "sources": [
    {
      "documentPath": "contract.pdf",
      "score": 9.2,
      "content": "Either party may terminate...",
      "heading": "Termination"
    }
  ],
  "time": 1523
}