Skip to content
Agentic WorkFlow

Local Q&A Model

Local Q&A Model

Section titled “Local Q&A Model”

An on-device extractive question-answering model you plug into the Model port of the Q&A Agent. With it connected, the agent answers locally in your browser by quoting the most relevant passage from its connected Knowledge — free, private, offline, and without made-up answers.

What it does

Section titled “What it does”

Exposes an installed extractive Q&A model (e.g. DistilBERT SQuAD) as a dependency. When the Q&A Agent runs with it, the agent:

  1. Retrieves the top passages from its connected Local Knowledge (the agent’s Knowledge documents (k) option, default 4).
  2. Runs the model over each passage (each capped to Max knowledge context length, default 4000 characters).
  3. Keeps the highest-scoring answer and tells you which document it came from.

Extractive models can only return text that is present in a passage, so the Q&A Agent must have a Knowledge connected in this mode.

When to use it

Section titled “When to use it”
  • Look up facts in your own documents without sending them or the question to a server.
  • You want answers that are verbatim quotes you can trust, not generated text.

For open-ended, synthesized answers, connect a chat model instead.

Inputs and settings

Section titled “Inputs and settings”
SettingNotes
ModelAn installed extractive Q&A model. Pick one, or download it from the picker. Cleared when a workflow is published, so each user picks their own.

Outputs

Section titled “Outputs”

Returns a Q&A model dependency for the Q&A Agent’s Model port. See the Q&A Agent for the output shape (response, answer, score, source).

Troubleshooting

Section titled “Troubleshooting”
  • ”… it needs a context: connect a Knowledge to the Q&A Agent.” — connect a Local Knowledge to the agent.
  • “Choose an on-device Q&A model in the Local Q&A Model node.” — the Model field is empty.
  • Empty answer — none of the retrieved passages contained an answer (a warning is logged). Raise Knowledge documents (k), rephrase the question, or check the knowledge store actually holds the information.
Section titled “Related pages”