Project 03
Grounded Retrieval-Augmented Assistant
The Unofficial Guide
The Unofficial Guide makes a small collection of student-review documents searchable through natural-language questions while keeping answers tied to retrieved evidence.
CodePath course project built from a provided starter repository.
Source documents
Overlapping chunks
Vector embeddings
Attributed answer
- 11
- source documents
- 35
- overlapping chunks
- 384
- embedding dimensions
Corpus and retrieval
Eleven source documents are split into 35 overlapping chunks. Sentence Transformers converts each chunk into a 384-dimensional embedding, and ChromaDB performs vector-similarity search against the question.
Grounded generation
The highest-similarity chunks are passed to Groq for generation under instructions to answer only from the supplied documents. Source attribution is appended programmatically from retrieved metadata rather than relying on the model to cite its sources correctly.
Testing the boundaries
The project tests retrieval quality, grounded answers, source display, and the refusal path used when the collection does not contain enough information. That makes out-of-scope behavior part of the product contract instead of an afterthought.
Project context
This was completed as a CodePath course project and remains visibly forked from the provided starter repository. The retrieval strategy, source collection, evaluation, and implementation documented here are the project contributions.