01 · Context
The problem
Institution-specific questions need institution-specific evidence. A general language model may know how to speak fluently without knowing the exact admission information a student is asking for.
02 · Product
How it works
The notebook builds an agent on IBM watsonx that uses a RAG utility tool to retrieve information from an MITAOE admission vector index, alongside a Google Search utility tool. The system prompt constrains the assistant to stay on-topic and keep responses short.
A dedicated RAG utility tool queries a vector index configured for MITAOE admission information.
The agent also exposes a Google Search utility through the watsonx toolkit.
Conversation state uses LangGraph-style memory via MemorySaver and thread IDs.
The system instructions explicitly constrain responses to the user's exact question.
Notebook code includes graph visualization for the agent workflow.
The repository remains intentionally small: a notebook, README and supporting PDF.
03 · Architecture
The system, broken down
Runtime
IBM watsonx API client supplies the project/space context and model runtime.
Agent
A LangChain ReAct agent combines the chat model with a set of tools and conversation memory.
Retrieval
The RAGQuery utility tool points to a specific vector index configured for MITAOE admission knowledge.
Web search
A GoogleSearch utility is added alongside retrieval when external lookup is needed.
Response policy
The agent prompt keeps responses short, focused on the question, and scoped to defined admission intents.
04 · Flow
A request through the system
Ask
A student enters a question such as fees, cutoff, scholarships or documents.
Choose a tool
The ReAct agent can use the admission RAG tool or Google Search to gather context.
Ground
Retrieved context is fed into the model before the response is produced.
Remember
A thread ID and MemorySaver keep the conversation state available for follow-ups.
Answer
The prompt asks the agent to return a concise, on-topic answer rather than dumping unrelated information.
05 · Engineering
Decisions worth noticing
Retrieval over pure prompting
The project explicitly adds a RAG tool so institution-specific answers can be grounded in external knowledge.
Tool-based agent
The notebook uses a ReAct agent instead of hard-coding a single retrieve-then-generate function.
Strict response policy
The prompt encodes topic and length constraints, which is useful for admissions assistance where users want direct answers.
Small, reproducible surface
Keeping the implementation inside a notebook makes the experiment easy to inspect and iterate on.
Scope note
This is the smallest of the four repositories, so the case study focuses on the implemented agent and retrieval workflow rather than inventing a web product around it.