Choose context
The user selects the application and the profile that matches their task.
How it works
Embedded authentication binds the principal to an application group and removes technical roles.
AI assistant for internal applications
An internal AI workspace that connects application knowledge, user roles and controlled tool execution.
Project overview
What the service does
An employee opens the assistant inside an internal product, asks a question about that product and gets an answer based on a recorded version of its source code. When the assistant wants to use a tool, the system checks what that tool is allowed to do and may ask the user to approve it.
Knowledge about internal applications is distributed across repositories and people. A useful assistant needs the right application context and clear limits on what it may do on behalf of a user.
Who uses it
Employees who support, operate or develop internal applications; business users can use business context, while technical users can work with source-code context.
Input
- A question and the selected application.
- The user profile and permitted capabilities.
- A synchronized source snapshot with its commit identifier.
Output
- An answer, a clarification request or a request to use a tool.
- A request the user can approve or deny before the tool runs.
- Saved conversation history and files produced by the workflow.
System requirements
- Business users see product information; developers can also use its source code.
- Each repository snapshot is tied to a specific revision.
- For long tasks, users can follow progress, answer questions and approve tool use.
- The chat opens inside the product the user is asking about.
01 / How the service works
Employee question → product context → permission check → answer
Process mapSolid line: normal flow · dashed line: retry or return
- INPUTEmployee question
Chosen product and user profile
ContinueProduct contextnew conversationConversation history - DATAProduct context
Source revision and code map
ContinueAssistant decision - DECISIONAssistant decision
Answer, ask, or request a tool
answerAnswerneeds detailsAsk for detailstool requestPermission check - DECISIONPermission check
Allowed, denied, or approval needed
allowedRun an approved toolapproval neededRequest approvaldeniedAction denied - SERVICERun an approved tool
Scoped access and execution limits
tool resultAnswer - RESULTAnswer
Saved reply for the conversation
saved replyConversation history - DATAConversation history
Question, events, answer and artifacts
- STEPAsk for details
Wait for the employee’s reply
- STEPRequest approval
The employee decides before execution
approvedRun an approved tool - RESULTAction denied
The tool is not run
02 / System architecture
Main components and their roles
Load knowledge
The assistant receives a versioned snapshot of the relevant repositories.
How it works
Background sync records the repository revision and builds a code map that helps the assistant find relevant files.
Record the run
The question and execution settings become a tracked task.
How it works
The scheduler does not start another run while one is active. PostgreSQL stores the current attempt and session state; Celery sends the task to a worker.
Plan the response
The worker can answer, request clarification or ask to use a tool.
How it works
The system checks the worker response before continuing the task. Tool use follows the selected run mode and timeout.
Check the action
The requested action is allowed, denied or sent for confirmation.
How it works
The policy checks the requested tool, run mode and conversation context. Access to local folders also requires browser permission for the selected folder.
Follow progress
The user sees progress and can return to a disconnected session.
How it works
Agent events are stored in order. If the connection drops, the frontend continues from the last received event.
Keep the result
The answer, context version and generated artifacts remain associated with the conversation.
How it works
The conversation keeps the assistant reply, repository revision and files created during the task.
03 / Implementation details
Engineering decisions
Permission before action
Tool requests pass through an explicit policy and approval state instead of executing solely on model output.
Versioned context
Repository revisions remain attached to the knowledge snapshot, so an answer can be related to the inspected source.
Resumable interaction
Persistent events and sequence cursors support continuity during long tasks and connection interruptions.
Software development for business
Have a project in mind?
Tell me what the application should do, how the team works today and which requirements or deadlines matter.