Unfolding the next page…
Unfolding the next page…
02 / Agents & knowledge systems
PlannedContext first. Action with permission.
Project type
Self-initiated portfolio project
Designed for
B2B SaaS · Support · Internal operations
Current stage
Specification & interface concept
The problem
Support teams jump between documentation, historical tickets and account tools to investigate a single issue. Answers lose context, and updates are easy to make without enough evidence.
The objective
Design a copilot that retrieves relevant knowledge, explains its sources, looks up approved account information and proposes changes for a human to approve.
Explore the concept
The product is planned. The interface below is an interactive concept with a fictional support issue and scripted responses. No live retrieval, model, account or ticketing system is connected. Measured evaluation results will be published after implementation.
COPILOT / SCRIPTED RESPONSE
The sample workspace reached its API request limit. Its connector is waiting for the next retry window.
Sources and account details are fictional. This is a walkthrough of the intended experience.
Retrieved context
“Connectors pause after the request budget is exhausted and resume at the next retry window.”
Fictional internal documentation · section 3.2
Planned architecture
A proposed architecture, to be validated during implementation.
Documentation & tickets → chunks + embeddings
Question + access scope → relevant sources
Sources + read-only tools → cited response
Proposed mutation → explicit human approval
Execution logs → retrieval & answer evaluation
Planned capabilities
Engineering decisions
Filter source access and tool scope before providing context to a model. A prompt asking the model to be careful is not an authorization boundary.
Return citations with responses and log retrieved passages, tool calls and decisions. Make unsupported answers and missing context visible.
Separate read-only investigation from proposed updates. Recheck authorization on execution and require a person to approve the intended change.
Evaluation plan
No results have been measured yet. These are the checks we plan to run and publish.
The product is planned. The interface below is an interactive concept with a fictional support issue and scripted responses. No live retrieval, model, account or ticketing system is connected. Measured evaluation results will be published after implementation.
Product repository · not published yet
Live product demo · not deployed yet
The next branch
Bring the messy workflow, the ambitious idea or the question you haven’t quite figured out.