Phase 05 · Embeddings & vector search Core

Build: semantic search end-to-end

The rite-of-passage build: ingest documents -> chunk -> embed -> index -> query - the complete retrieval pipeline that RAG will sit on.

In plain terms

Everything so far - embeddings, chunks, vector stores - are parts on a workbench. This node is assembling the watch: a search box over YOUR documents that finds meaning, not keywords. Type 'money back' and get the refund policy, which never contains those words.

Why it matters

Concepts don't consolidate until wired together. Every real failure mode - bad chunks, truncation, junk matches with high scores - only appears when the pipeline runs end to end. This build is also literally the retrieval half of RAG; do it well and Phase 05 becomes easy.

How it works

Pipeline: (1) load documents (start with 20+ markdown/text files), (2) chunk by structure ~500 tokens, (3) embed each chunk, (4) upsert into Chroma with metadata, (5) at query time: embed the query, fetch top-5, display with source + score. Then the important part: run 20 realistic queries and study the failures - that study IS the skill.

When you use it

Now. Before touching RAG frameworks - you must know what they're wrapping.

Common mistakes

  • Reaching for LangChain before building it raw once - the framework hides exactly the parts you need to understand.
  • Testing with 3 easy queries and declaring victory; failures live in queries 10-20.
  • Skipping the score display, so you can't see that your 'best match' is a mediocre 0.4.

Best practices

  • Log every query with its top-5 results and scores; review weekly like a detective.
  • Add a tiny golden set: 10 queries with known-correct chunks, and % correct-in-top-5 as your score (-> evals).
  • Show sources in results from the very first version.

Try it yourself

Ship the pipeline above over your own notes or a project's docs.

Project: Phase project - search your own life. Semantic search over something you actually own: your notes, bookmarks, chat exports or a favorite book series. Requirements: 100+ chunks, metadata filters, scores + sources displayed, a golden set of 10 test queries with measured hit-rate, and a README explaining your chunking choice. You'll use this exact index again in Phase 05.

Resources

Related topics

Chunking strategies Splitting documents into right-sized pieces before embedding - the humble step that quietly decides retrieval quality more than any model choice. Vector databases Databases purpose-built to store millions of embeddings and answer 'find the K nearest vectors to this one' in milliseconds. The RAG pipeline, end to end Wiring ingestion and query-time retrieval into one system, where overall quality is the product of every stage's quality. What are embeddings An embedding turns text into a list of numbers (a vector) positioned so that similar meanings land close together in space. What is RAG & why it exists RAG = retrieve relevant documents for the user's question, paste them into the prompt, and instruct the model to answer only from them.