Advanced Retrieval for AI with Chroma
by DeepLearning.AI
What to do when plain vector search returns the wrong chunks — query expansion, re-ranking and embedding adaptors.
Overview
Taught by Anton Troynikov, co-founder of the open-source vector database Chroma, the course runs 1 hour 2 minutes across seven lessons with five runnable code examples and a graded quiz. The sequence is Introduction; Overview of embeddings-based retrieval; Pitfalls of retrieval — when simple vector search fails; Query Expansion; Cross-encoder re-ranking; Embedding adaptors; and Other Techniques. The framing is diagnostic rather than promotional: lesson three is dedicated to recognising when a retrieval pipeline is returning semantically near but practically useless results, which is the failure most RAG tutorials never mention. The repairs that follow are each a distinct lever. Query expansion uses an LLM to generate additional related queries or a hypothetical answer document, widening recall beyond the literal user phrasing. Cross-encoder re-ranking scores query and candidate document jointly rather than comparing pre-computed vectors, then reorders results by that sharper similarity measure. Embedding adaptors go further, training a small model on user relevance feedback that transforms embedding values to suit your specific corpus. Chroma itself is Apache 2.0 licensed with roughly 29,000 GitHub stars and a weekly release cadence, so the underlying library the notebooks use is actively maintained. Prerequisite is intermediate Python; enrollment is free and short courses issue a completion email rather than a formal certificate.
At a Glance
- Topic
- RAG
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 1 hour 2 minutes, self-paced (7 lessons, 5 code examples, 1 graded quiz)
- Provider
- DeepLearning.AI
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Recognise when a retrieval pipeline is returning semantically close but practically irrelevant results
- ✓Diagnose the specific failure modes of simple embeddings-based vector search on real corpora
- ✓Use an LLM to expand a user query into related queries or a hypothetical answer document
- ✓Apply cross-encoder re-ranking to reorder retrieved candidates by joint query-document scoring
- ✓Train an embedding adaptor from user feedback to reshape embeddings toward your domain
- ✓Project and visualise an embedding space to see why particular retrievals went wrong
- ✓Choose between expansion, re-ranking and adaptation based on which failure you actually have
Highlights
- •Leads with retrieval failure rather than the happy path — rare and genuinely useful framing
- •Taught by Chroma's co-founder, so the vector-store internals come first-hand
- •Embedding adaptors are barely covered anywhere else at this length and accessibility
- •Five of seven lessons carry runnable code, so it is closer to a workshop than a lecture
- •Built on Chroma, which is Apache 2.0 with ~29k GitHub stars and weekly releases — no vendor lock-in to evaluate
Who It’s For
Best For
- ✓Engineers whose RAG prototype demos well but retrieves poorly on real queries
- ✓Teams tuning retrieval quality before reaching for a larger or more expensive model
- ✓Developers who already know basic RAG and want the next layer of technique
- ✓ML practitioners evaluating whether to fine-tune embeddings for a domain corpus
Prerequisites
- •Intermediate Python — the course assumes you can read and modify notebook code confidently
- •Working knowledge of embeddings and basic retrieval-augmented generation
- •An OpenAI API key to run the query-expansion and generation examples
FAQ
What is Advanced Retrieval for AI with Chroma?
A short course about the failure modes of naive RAG. It starts by showing where simple embeddings-based retrieval breaks down, then works through three concrete repairs: expanding queries with an LLM, re-ranking retrieved results with a cross-encoder, and training an embedding adaptor to reshape the embedding space toward your domain. For engineers whose RAG prototype works in the demo and disappoints in production.
Is Advanced Retrieval for AI with Chroma free?
Advanced Retrieval for AI with Chroma is free to access.
What level is Advanced Retrieval for AI with Chroma for?
Advanced Retrieval for AI with Chroma is aimed at a intermediate audience. Recommended background: Intermediate Python — the course assumes you can read and modify notebook code confidently, Working knowledge of embeddings and basic retrieval-augmented generation, An OpenAI API key to run the query-expansion and generation examples.
How long does Advanced Retrieval for AI with Chroma take?
Expect roughly 1 hour 2 minutes, self-paced (7 lessons, 5 code examples, 1 graded quiz). Most learners work through it at their own pace.
What will I learn from Advanced Retrieval for AI with Chroma?
You'll learn: Recognise when a retrieval pipeline is returning semantically close but practically irrelevant results; Diagnose the specific failure modes of simple embeddings-based vector search on real corpora; Use an LLM to expand a user query into related queries or a hypothetical answer document; Apply cross-encoder re-ranking to reorder retrieved candidates by joint query-document scoring; Train an embedding adaptor from user feedback to reshape embeddings toward your domain; Project and visualise an embedding space to see why particular retrievals went wrong; Choose between expansion, re-ranking and adaptation based on which failure you actually have.
Topics
Sources
This page was written from 4 sources, 3 on domains other than deeplearning.ai.