A single retrieval chat does one thing per question: search, then answer. Wyregraf Pro's Agents and Networks let a chat do several things in sequence — rewrite a vague question into a searchable one, retrieve, consult conversation history, and shape the reply — each step handled by a dedicated agent.
They're an advanced feature; a plain retrieval chat is the right starting point for most sites.
Agents #
An Agent is one reusable unit of AI work: a tool, plus the prompts and profiles it runs with. Go to Wyregraf → Agents.
- Add New creates an agent.
- Each agent in the list can be used by any network.
- Agent Type sets how the agent runs.
- The editor's tabs hold the rest: Tool, Prompts, and Profiles.
The Tool tab is what the agent actually does:
- Pick one of the built-in tools:
- Retrieval — search your embeddings and answer from the matches.
- Query Processor — rewrite the visitor's raw question into a form that retrieves well.
- History — answer from the conversation so far.
- Lever — adjust the response along configured dimensions such as tone, clarity, or formality.
The Prompts tab sets the agent's system prompt and prompt (manage reusable ones under Wyregraf → All Prompts), and Profiles picks its embeddings/LLM profiles.
Networks #
A Network wires agents into a pipeline. Go to Wyregraf → Networks.
- The Processing Agent receives each incoming question first — typically a Query Processor that decides and prepares what happens next.
Below it, the Agents list is the pipeline itself. Each row pairs an agent with its prompt, a Fallback Agent to run if it comes up empty (for example: retrieval first, fall back to history), and a Response Evaluation that checks the step's output before it moves on.
Putting a network behind a chat #
On a chat's Main tab, set Chat Type to Agent Network and pick the network. The visitor experience stays the familiar chat widget from Displaying the Chat — the network only changes how the answer is produced.
