Free project guide / Intermediate

Build a source-aware document search

Start with retrieval you can inspect before adding generated answers.

3–5 practice sessions · suggested effort

What you'll make

A small search prototype that returns source passages and exposes where it fails.

Choose a small collection

Use documents you have permission to use. Keep a source title and document ID. Write ten questions, including two the collection cannot answer.

Checkpoint: You can identify the expected source for each answerable question.

Split and label passages

Divide text at paragraph boundaries. Store document ID, passage ID, title, and text together. Inspect the splits manually.

Checkpoint: No passage has lost the information needed to locate its source.

Create a retrieval baseline

Tokenize the question and rank passages by matching terms. Return the highest-ranked passages with their titles and IDs. This is a baseline, not semantic search.

Checkpoint: You can explain why each result was returned.

Measure before improving

For each test question, record whether a relevant passage appears among the first three results. Inspect queries that fail. Consider embeddings only after saving the baseline.

Checkpoint: You have a small results table rather than a general claim that the system works.

Optionally add generated answers

Give a model only the retrieved passages and ask for source IDs. Verify every answer against its source. Keep API keys in local environment variables and set a spending limit.

Checkpoint: Unsupported questions are handled explicitly, and citations actually support the response.

Before you call it finished

Use this checklist to review your project. Sign in to save your checklist and continue from your dashboard.

Document one thing that failed, how you investigated it, and what you would improve next.

Bring your questions to a conversation.

Explain your goal, what you tried, and where you got stuck. Check the available mentoring packages for scope and price.

Explore mentoring