A client asks for a progress update. The useful answer depends on a brief, the current project status and recent decisions, often held in different systems. Hookie can bring these inputs into a Workspace Brain, where a workflow or connected agent retrieves the relevant records and relationships before drafting the update.
This pipeline suits small businesses coordinating sales and delivery, freelancers managing client projects, and business AI agents preparing support responses or operational briefings. Workspace Brain and database sources require Pro or Team.
1. Bring in the updates
Choose the intake that fits each source:
- Webhooks: receive form submissions, customer events and project updates at a dedicated endpoint. Configure supported provider signatures and map the payload into a dataset.
- WebSocket listeners: connect to an external feed, filter and map its messages, and use the resulting records to trigger workflows.
- PostgreSQL sources: poll a table for new rows using a unique, increasing cursor. Each stored row enters the source’s dataset and can trigger the same downstream processing.
Use the webhook quickstart, workflow guide and database source guide for setup. Database sources are scheduled polling; the polling interval determines when new rows are read.
2. Map the records into useful tables
Create Clients, Projects and Briefs in Data workspace. Choose fields such as client reference, project status, summary and decision date. Connect each source dataset and map its fields into that structure. Preview a sample before saving the mapping.
Use stable references to identify the same client or project across inputs. Configure workflow table upserts when you want to update an existing record; a new event is not automatically the same entity as an older one. Add AI enrichment where it is useful, then review the output before treating it as an approved business fact.
This repeatable mapping reduces manual preparation and makes missing or inconsistent fields easier to catch. Review incomplete source records before using them as agent context.
3. Connect the knowledge and choose its scope
Table records are reflected in the graph. Connect a client to a project and the project to its brief with explicit relationships, created through the interface, a Knowledge graph workflow step or the API.
For example, define the relationships client → project → brief. The client becomes a starting point for retrieving its project context. You control the meaning and direction of each connection.
Include only the tables and records suited to the Brain. Excluding a record keeps it in its source table but removes it from graph discovery and agent context. Use these controls to leave unrelated work out of the context pipeline; access permissions still determine who can read workspace data.
4. Fetch context before calling the agent
Grant the workflow’s project Read graph context access. Add a Knowledge graph → Fetch agent context step, then an AI or Agent step. Pass the fetched result into the agent’s input.
Choose the starting node, the relationship direction and types, and the fields needed for the task. Limit how far traversal can go, how many records it returns and the serialized response size. A briefing may need summaries and statuses; it may not need every raw event or every property on those records.
The Workspace Brain guide includes the workflow configuration and the query_graph_context MCP tool for connected agents. Context responses expose truncation flags, so an agent can see when the configured limit omitted information and request more when needed.
5. Review the answer and keep the knowledge current
Have the agent draft the client update from the retrieved facts, identify missing information and leave the final review to a person. Inspect the source records and workflow execution when something looks wrong. Revision checks preserve conflicting agent edits as proposals for review instead of silently replacing a newer human edit.
Agent and workflow graph changes are queued, usually consolidating after about 15 seconds. A receipt confirms a queued change; wait for it to be applied before another task depends on that new context. Backlog or outages can delay application.
Why this uses context more efficiently
Retrieving selected fields from related records can produce a smaller prompt than sending full records alongside unrelated client history. That can reduce input tokens and gives the model less text to process. Smaller responses can also reduce transfer and model processing time; measure the result with your own data, task and model.
The accuracy benefit comes from consistent mappings, explicit relationships, source provenance and human curation. Review the retrieved context as well as the generated answer, and keep records up to date.
Explore knowledge graphs · Compare plans · Connect an AI agent