← Back to Paths
[PLACEHOLDER hero banner]
RAG & LLM Application Engineering
Ship production RAG pipelines that retrieve accurately and answer reliably at scale.
CREATED BY
D
Dev R. [PLACEHOLDER] ★ 4.8
Senior Data Engineer at StreamBase | 10+ years of experience
About this Path
For engineers building real products on top of LLMs. Goes beyond hello-world chains to cover retrieval architecture, embedding strategy, evaluation frameworks, and latency-cost trade-offs. You will build an end-to-end RAG system with chunking, hybrid search, reranking, and an automated eval harness.
Path Overview
Advanced LevelCertificate of CompletionAbout 48 hours to completeEnglish language16+ curated videosLearn online at your own pace5 modules with resourcesGamified & interactive
Path Curriculum
Naive RAG vs Advanced RAG vs Modular RAG
Understand trade-offs before choosing architecture for your data and query profile.
Chunking Strategies — Fixed, Semantic, and Recursive
Impact of chunk size and overlap on recall; parent-document retriever pattern.
Embedding Model Selection and Trade-offs
OpenAI text-embedding-3 vs BGE vs E5; dimensionality, cost, and domain fit.
Vector Store Internals — HNSW, IVF, PQ
Index types in Pinecone, Weaviate, and pgvector; recall vs latency trade-offs.
BM25 Sparse Retrieval with Elasticsearch
Tokenisation, IDF tuning, and integrating keyword results alongside vector hits.
Reciprocal Rank Fusion for Score Merging
Combine dense and sparse rankings without calibrating heterogeneous score spaces.
Cross-Encoder Reranking with Cohere Rerank
Re-score top-k passages for precision; latency budget and k-selection heuristics.
Contextual Compression and Document Filtering
Strip irrelevant sentences from retrieved chunks before injecting into context window.
Structured System Prompts for Grounded Answers
Instruct the model to cite source chunks and refuse when context is insufficient.
Chain-of-Thought and Step-Back Prompting
Improve multi-hop reasoning by reformulating queries before retrieval.
Conversation History Management and Token Budgeting
Rolling window, summarisation, and selective history injection strategies.
RAGAS Metrics — Faithfulness, Answer Relevancy, Context Recall
Run automated eval pipelines on a golden QA dataset after every pipeline change.
LLM-as-Judge Patterns with GPT-4o
Build pairwise and reference-free judges; calibrate against human labels.
Regression Testing and Eval CI Integration
Gate pull requests on RAGAS score thresholds using GitHub Actions.
Hallucination Detection with NLI Models
Classify generated claims against source chunks using a entailment classifier.
Semantic Caching with Redis and GPTCache
Cache embedding lookups and LLM responses for repeated or near-duplicate queries.
Streaming Responses with FastAPI and Server-Sent Events
Deliver tokens progressively to reduce perceived latency in chat interfaces.
Observability — Langfuse Traces and Token Spend Dashboards
Instrument every retrieval and generation step; alert on cost and latency spikes.
Kubernetes Deployment with GPU Node Pools
Deploy embedding and reranker inference containers with resource requests and limits.
What you'll learn
- ✓Design a retrieval pipeline with chunking strategies, embedding model selection, and vector store trade-offs.
- ✓Implement hybrid search combining dense vector similarity and BM25 sparse retrieval with Reciprocal Rank Fusion.
- ✓Fine-tune retrieval quality using cross-encoder rerankers and contextual compression to cut hallucination rates.
- ✓Evaluate RAG quality systematically using RAGAS metrics — faithfulness, answer relevancy, and context recall.
- ✓Reduce latency and cost with semantic caching, batched embedding, and streaming token delivery.
- ✓Deploy a RAG API on Kubernetes with observability hooks for retrieval latency, LLM token spend, and error rates.