Skip to content
LeetLLM
Tracks
Intermediate22 lessons

RAG and Search Systems

Help someone find an answer in documents they do not have time to search by hand, with sources they can check.

Builders working on search, support bots, internal knowledge assistants, and document QA.

You can design a retrieval pipeline, debug faithfulness failures, and choose vector, lexical, and graph retrieval pieces deliberately.

  1. 1Static to Contextual EmbeddingsThe same token ID starts from one table row, while invoice and phone clues can move each occurrence toward a different meaning.Core LLM FoundationsMedium26 min
  2. 2Sentence Embeddings & Contrastive LossTrain vectors so the right passage outranks a lookalike, then check that ranking against hard negatives, reranking cost, and a smaller index.Transformer Deep DivesHard67 min
  3. 3Embedding Similarity & QuantizationA longer vector can outrank the right passage, and a coarser code can change the ranking again. Measure which contract moved.Transformer Deep DivesHard57 min
  4. 4Dimensionality Reduction for EmbeddingsShortening a policy embedding saves memory and can drop the clause that answers the question. Measure recall before trusting PCA, a projection, or quantization.Applied LLM EngineeringMedium40 min
  5. 5File Ingestion for AIThe assistant cited a 14-day archive rule because the 15-minute rollback line never survived extraction.Core LLM FoundationsMedium22 min
  6. 6Chunking StrategiesTurn clean documents into retrieval units that preserve answers, citations, and measurable search quality.Core LLM FoundationsMedium22 min
  7. 7Core Retrieval AlgorithmsSearch can return the API-key quota page instead of the rotation steps. Lexical scores, dense vectors, fusion, and approximate indexes each fail that question differently.ML Algorithms & EvaluationMedium32 min
  8. 8Production RAG PipelinesMake RAG answer from the right current policy for the right caller, and abstain when its evidence falls short.Applied LLM EngineeringMedium37 min
  9. 9Vector DB Internals: HNSW & IVFFind out how an index can miss the rollback runbook, then tune HNSW, IVF, and compression against recall, latency, and memory.Advanced Agents & RetrievalHard57 min
  10. 10Hybrid Search: Dense + SparseHelp Luna find the right EU key-rotation policy by code or paraphrase, while keeping expired and restricted rules out of her results.Applied LLM EngineeringMedium34 min
  11. 11Reranking and Cross-Encoders for RAGMove the right policy into the prompt. Learn how cross-encoders reorder retrieved passages, and where recall, permissions, and text limits still matter.Applied LLM EngineeringMedium36 min
  12. 12RAG Evaluation for Reliable AnswersTest whether a RAG answer is supported, cites the right source, and covers every required point. Catch a convincing answer that still gets the policy wrong.Applied LLM EngineeringMedium31 min
  13. 13LLM-as-a-Judge EvaluationTwo rollback replies both follow DEP-27, but a string match can't say which one an on-call engineer can use.Applied LLM EngineeringMedium26 min
  14. 14Hallucination Detection & MitigationAn incident draft promises recovery by 14:30 UTC, but the admitted runbook has no such time. The gate has to decide what may still be served.Applied LLM EngineeringMedium31 min
  15. 15LLM Observability & MonitoringSix responses return HTTP 200 and still include two unsupported approvals. Latency, cost, and traces matter only once that claim is visible.Applied LLM EngineeringMedium32 min
  16. 16Advanced RAG: HyDE & Self-RAGA fast index can still miss what a user means. Learn when query rewriting, HyDE, Self-RAG, and CRAG help, then test the trade-off on a docs assistant.Advanced Agents & RetrievalHard55 min
  17. 17GraphRAG & Knowledge GraphsTop-k may explain one incident and miss the fleet pattern. Source-linked reports support broader answers when coverage, identity, and grants hold.Advanced Agents & RetrievalHard57 min
  18. 18RAG Security & Access ControlStop a signing-key query from exposing a restricted runbook: authorize before export, isolate tenants, and audit the reject path.Advanced Agents & RetrievalHard58 min
  19. 19Model Versioning & DeploymentAn engineer reverts the prompt in Git, but the image and the evidence have already moved. The release ID has to name the combination that is actually answering.Applied LLM EngineeringMedium33 min
  20. 20Semantic Caching & Cost OptimizationReuse a policy answer across paraphrases only after release, access, and freshness boundaries hold, and only when quality, cost, and tail latency still justify the hit.Applied LLM EngineeringMedium33 min
  21. 21Capstone: Document QAA published policy and a private note can both be retrieved. The lookup has to keep the note out and still cite only a passage that supports the question.Portfolio CapstonesHard54 min
  22. 22LLM-Powered Search EngineHelp a developer find the current parser rule without exposing another team's roadmap. Design a search path that preserves its exception and shows only checked claims.System Design CapstonesHard37 min