Cheat sheet template
A full one-page cheat sheet on hybrid RAG, laid out in three columns: what it is, an overview flow, the architecture and more. Every block is an ordinary shape on the canvas, so you can rewrite it for any topic.
Good for
- A one-page explainer for a concept or a tool
- Revision notes before an interview or an exam
- A shareable summary for a talk or a course
How to use it
- Press Use this template. It opens on the canvas in your browser, free, with no sign-up.
- Double-click any shape to rename it, or press / and type a line such as
Client -> API -> DBto add to it. - Share it as a link, present it with Shift+P, or export PNG, SVG or PDF.
In this diagram
- Hybrid RAG
- Retrieval-augmented generation that runs several retrievers in parallel, then merges and reranks before the LLM answers.
- ① What is Hybrid RAG?
- Combines multiple retrieval strategies (e.g. Dense + Sparse + Others)
- Leverages the strengths of each method
- Improves recall, relevance & robustness
- Delivers higher quality context to the LLM
- ② Overview
- User query
- Hybrid retriever
- Retrieve
- Merged top-k docs
- Merge
- Reranker (optional)
- Rerank
- LLM (generator)
- Generate
- Answer
- ③ Architecture
- Query processing
- Understand and transform the query (if needed)
- Dense (vector DB)
- Sparse (BM25)
- Other (e.g. graph)
- Retrieve relevant documents using multiple methods in parallel
- Merge & deduplicate candidates
- Candidate pool: combine results and remove duplicates
- Rerank top-N candidates
- Reranker (optional): improve ordering based on relevance
- Build context for LLM (stuff / map-reduce)
- Context builder: prepare final context for the LLM
- Generate answer
- LLM (generator): generate response using the context
- Answer (response)
- Return final answer to the user
- ④ How it works
- User provides a query
- Query is processed / rewritten (optional)
- Multiple retrievers search in parallel
- Results are merged & duplicates removed
- Candidates are reranked for relevance
- Top-k documents are selected
- Context is built and passed to the LLM
- LLM generates the answer
- Response is returned to the user
- ⑤ Retrieval strategies
- ⑥ Example
- “How does RAG improve LLM accuracy?”
- Parallel retrieval
- Dense, sparse and graph retrievers search at the same time
- Merge + deduplicate
- Top 50 candidates
- Top 10
- LLM context
- LLM answer
- ⑦ Benefits & challenges
- Benefits
- ✓ Higher recall & better relevance
- ✓ Robust across query types
- ✓ Handles semantic + keyword queries
- ✓ Improves answer quality
- Challenges
- ✗ More complex to build
- ✗ Higher latency & cost
- ✗ Needs good merging & reranking
- ✗ Harder to tune & optimize
- ⑧ Hybrid RAG variants
- Dense + Sparse
- Combine semantic (dense) and keyword (sparse) search
- Most common hybrid approach
- Multi-stage hybrid
- Use one retriever first, then another for refinement
- Improves precision
- Router-based hybrid
- Route the query to the best retriever by intent or type
- Dynamic & adaptive
- Ensemble hybrid
- Combine multiple retrievers and rankers with weighted scoring
- Highest robustness
- ⑨ Best practices