LangChain
Build · Data
Connect model calls to tools, document loaders and retrieval components.
All capabilities · Retrieval & knowledge systems
Build a grounded RAG application with citations
Retrieve → rerank → generate grounded answers with citations over a private corpus.
Explore 5 tools for this projectMarket relevance
Which roles ask for this — and how often
Share of job postings in India, per role, that name this capability.
What employers mean
You should be able to…
- Build a retrieve → rerank → generate pipeline that answers questions grounded in a private corpus
- Return citations/source references alongside generated answers
- Handle 'I don't know' gracefully when the corpus doesn't contain an answer, instead of hallucinating
- Tune retrieval (top-k, chunk size, reranking) to improve answer quality
- Support multi-turn Q&A where follow-up questions still retrieve correctly
- Add a reranking step to improve precision after initial vector retrieval
- Deploy the RAG system behind an API or simple UI for others to query
Needs first: Generate embeddings and run vector search, Design and version prompts systematically
Learn — free, link-checked
The few resources that matter
Course · intermediate · 90 min · deeplearning.ai
LangChain: Chat with Your Data
End-to-end walkthrough from document loading to a working conversational retrieval chain. — DeepLearning.AI (Harrison Chase, Andrew Ng)
Build from · intermediate · 90 min · github.com
Anthropic Cookbook
Official recipes including contextual retrieval and RAG evaluation patterns tuned specifically for Claude. — Anthropic
Build from · intermediate · 120 min · github.com
RAG From Scratch
Builds every RAG stage (query translation, routing, retrieval, generation) from first principles instead of one black-box chain. — LangChain (langchain-ai)
Read · advanced · 40 min · huggingface.co
Advanced RAG on Hugging Face Documentation using LangChain
Covers reranking and query optimization plus a worked evaluation section, bridging pipeline-building and eval in one place. — Hugging Face
Course · advanced · 90 min · deeplearning.ai
Building and Evaluating Advanced RAG Applications
Pairs advanced retrieval techniques (sentence-window, auto-merging) directly with the eval metrics used to justify them. — DeepLearning.AI (Jerry Liu, Andrew Ng)
Tools for practice
Choose a tool for the job
Start with one tool for each part of your project. You don’t need to learn them all.
5 tools to explore
LlamaIndex
Build · Data
Connect documents to retrieval and model workflows with explicit data handling.
pgvector
Data · Build
Store embeddings in PostgreSQL and compare similarity-search results.
Chroma
Data · Build
Index embeddings and metadata, then retrieve relevant records for a query.
Claude API
Build · Test
Build model-backed features with messages, tool use and responses you can evaluate.
Practices & references
- Reranking
- Citations
- Failure analysis
Practice
Grounded RAG assistant over Indian labour-law documents
Build a RAG assistant over India's labour codes and related public labour-law documents, answering questions like 'how many days of casual leave am I entitled to?' with citations down to the section. Implement retrieve, rerank, then generate, and make the assistant say the corpus does not cover it rather than invent an answer. Expose it as a small API or UI that handles multi-turn follow-ups, and score it on a question set you label by hand.
Start from
Milestones
- Ingest the labour-law PDFs with section metadata and get vector retrieval working · ~4h
- Wire retrieve into generate with inline citations back to section and page · ~3h
- Add the reranker and compare retrieved chunks before and after · ~3h
- Add out-of-corpus refusal and multi-turn follow-up handling · ~3h
- Hand-label 10 questions, score the answers, deploy or record a demo · ~2h
Done when
- Answers include inline citations pointing to the specific source document and section/page
- A reranking step measurably improves top-result relevance over vector-only retrieval, shown with before/after examples
- At least 5 out-of-corpus questions are correctly answered with 'not covered' instead of a hallucinated answer
- A simple API or UI lets a user ask multi-turn follow-up questions that retrieve correctly
Prove it
Evidence a recruiter can check
- Ten answers with inline citations, each traceable to the section of the code it quotes
- The same query's top chunks before and after reranking, placed side by side
- Five out-of-corpus questions answered with an explicit 'not covered', pasted verbatim
- A hand-labelled score (e.g. 8 of 10 correct-with-citation) with the labelled question set committed
- A hosted link or demo recording where someone else can ask a question of their own
Signal it
Built a grounded RAG assistant over India's labour codes - retrieve, rerank and generate with section-level citations - scored on a hand-labelled question set and refusing out-of-corpus questions instead of hallucinating.
Interview
Questions you'll get asked
- Walk me through the architecture of a RAG pipeline you've built, retrieval to generation.
- How do you get an LLM to say 'I don't know' instead of hallucinating when the answer isn't in the retrieved context?
- What's the role of a reranker in a RAG pipeline, and when is it worth the added latency?
- How do you generate accurate citations that point back to the exact source chunk?
- How would you handle a follow-up question in a multi-turn RAG conversation where the retrieval query needs the prior turn's context?
- How do you decide top-k for retrieval, and what happens if it's too low or too high?
- How would you extend a RAG system to work well over a mix of English and Hindi documents?