Skip to main content
Every message runs through the same pipeline. Each stage is a guard or a decision before any expensive work happens, which is what keeps MIND predictable and cheap.

​
The request envelope

{
  "action": "free-chat",
  "body": "How do I submit my assignment?",
  "context": {
    "conversation_id": "1f0c…",
    "page": { "title": "Assignment 2", "course_code": "CSC201" }
  }
}
  • action — the quick-action hint, or free-chat when the student just typed.
  • body — the student’s message.
  • context.conversation_id — the thread to continue. Omit it on the first message; the response returns one to reuse.
  • context.page — what the student is looking at, used as a routing/answering hint.
Identity is taken only from the verified JWT — tenant, student, and role. Any user object in the request body is ignored. The client cannot assert who it is.

​
The pipeline

1

Context card

Identity is read from the JWT into a context card (student, tenant, role, and any programme scope present in the claims). Academic signals (level, performance) come from a read-only profile store and are simply absent until that store is wired — never guessed.
2

Daily quota

A per-student daily counter. When a cap is configured and exceeded, the turn ends immediately with a message event — no model call. Unlimited by default until a cap is set.
3

Input guardrails

Run before any model call: an assessment lock (MIND pauses during an active assessment) and a light prompt-injection check. A block ends the turn with a message event.
4

Routing

A cheap, fast model classifies the message into a lane. An explicit quick action is trusted as-is (no model call); free-chat is classified and may be upgraded — e.g. “How do I submit my assignment?” becomes submission-guide. The resolved lane is emitted as the meta event.
5

Conversational graph

The message enters a stateful graph that loads the conversation history, composes the system prompt and context, and generates the reply. History is persisted per conversation (see Memory).
6

Streaming

Tokens stream out as delta events the moment the model produces them.
7

Output guardrails

The complete answer is checked (PII redaction now; an answer-leak judge later). Any correction is surfaced on the done event as revised.

​
Memory and resumption

The conversation lives in a checkpoint keyed by conversation_id. Each turn appends the student’s message and MIND’s reply; the next request with the same conversation_id automatically loads the full history — no transcript needs to be replayed by the client. This makes resumption a property of the application, not the transport. If the network drops mid-stream, the client simply re-issues the request with the same conversation_id and the thread continues. To start fresh, omit conversation_id and a new one is returned.
Only a recent window of the history is sent to the model each turn, so cost stays bounded even on long conversations. The full history is always retained in the checkpoint.

​
The core services

The pipeline is assembled from a small set of channel-agnostic services. Each one has a single implementation today and a clean seam to swap later, so the chat surface never changes when an internal provider does.
ServiceTodayDesigned to become
GatewayGoogle Gemini, with separate router and reasoning tiersA self-hosted LiteLLM gateway
PromptsManaged in Langfuse, with an in-repo fallback—
ContextIdentity from the JWT+ academic signals from the profile store
GuardrailsAssessment lock, injection, PII+ LLM-judged answer-leak, full PII
QuotaDaily per-student counterShared/Redis-backed limits
RetrievalSingle payload-partitioned collection for course materialsGrounded answers as a tool

​
Prompts are versioned

System and router prompts are managed in Langfuse and resolved at request time, with an in-repo copy as a fallback if Langfuse is unavailable. Every prompt carries a version stamp, so a given reply can always be traced back to the exact prompt that produced it. Editing a prompt in Langfuse changes behaviour without a deploy — which is the primary way MIND’s tone and guardrails are tightened.