Context
Our client operates an education platform with a substantial library of learning content. Learners could search it, but searching is not the same as getting an answer.
Problem
A learner asking “how does this concept apply here?” does not want ten documents ranked by keyword. They want the answer, drawn from the material they are actually studying.
Putting a general model in front of that library does not solve it either. The model answers from its training data, which means answers that sound right, do not match the course, and occasionally contradict it. On an education platform that is worse than no feature at all — the whole product depends on the content being authoritative.
The requirement was that answers come from the client’s own content, and that the retrieval step find the right passage before the model writes anything.
Approach
Retrieval quality is the entire game in a RAG system. If the wrong passage is retrieved, no amount of prompt engineering saves the answer.
NEEDS INPUT on the specifics, which are exactly what a technical reader is looking for:
- How content was chunked, and how chunk boundaries were chosen for learning material.
- Embedding model and vector store.
- Whether retrieval is semantic only, or hybrid with keyword search.
- Whether re-ranking was used.
- How relevance was measured before and after — this is the number the Results section needs.
What we built
A retrieval-augmented generation pipeline over the platform’s content sources, improving answer relevance and contextual retrieval for learners.
Results
NEEDS INPUT. Improvement in answer relevance, retrieval accuracy on an evaluation set, or reduction in learner support queries.
Related
Serving a 72B Open-Source Model Under Concurrent Load · Computer Vision for Classroom Engagement · AI Engineering