[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M72XSNFV86ENBVW4KF4DM720":3,"skill-same:01M72XSNFV86ENBVW4KF4DM720":45},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":28,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M72XSNFV86ENBVW4KF4DM720","rag-implementation",null,"Build Retrieval-Augmented Generation (RAG) systems for LLM applications with vector databases and semantic search. Use when implementing knowledge-grounded AI, building document Q&A systems, or integrating LLMs with external knowledge bases.","steph-frtech\u002Fclaude-code-up","steph-frtech","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F198466323?v=4","01M72XD213FVQA2QXRGTJFKHE1","01M72XD227AANJ993SQTNZB1CK",1,"MIT","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fmarketplaces\u002Fwshobson\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Frag-implementation\u002FSKILL.md",[],[],"2026-11-10 22:39:42",false,"---\nname: rag-implementation\ndescription: Build Retrieval-Augmented Generation (RAG) systems for LLM applications with vector databases and semantic search. Use when implementing knowledge-grounded AI, building document Q&A systems, or integrating LLMs with external knowledge bases.\n---\n\n# RAG Implementation\n\nMaster Retrieval-Augmented Generation (RAG) to build LLM applications that provide accurate, grounded responses using external knowledge sources.\n\n## When to Use This Skill\n\n- Building Q&A systems over proprietary documents\n- Creating chatbots with current, factual information\n- Implementing semantic search with natural language queries\n- Reducing hallucinations with grounded responses\n- Enabling LLMs to access domain-specific knowledge\n- Building documentation assistants\n- Creating research tools with source citation\n\n## Core Components\n\n### 1. Vector Databases\n\n**Purpose**: Store and retrieve document embeddings efficiently\n\n**Options:**\n\n- **Pinecone**: Managed, scalable, serverless\n- **Weaviate**: Open-source, hybrid search, GraphQL\n- **Milvus**: High performance, on-premise\n- **Chroma**: Lightweight, easy to use, local development\n- **Qdrant**: Fast, filtered search, Rust-based\n- **pgvector**: PostgreSQL extension, SQL integration\n\n### 2. Embeddings\n\n**Purpose**: Convert text to numerical vectors for similarity search\n\n**Models (2026):**\n| Model | Dimensions | Best For |\n|-------|------------|----------|\n| **voyage-3-large** | 1024 | Claude apps (Anthropic recommended) |\n| **voyage-code-3** | 1024 | Code search |\n| **text-embedding-3-large** | 3072 | OpenAI apps, high accuracy |\n| **text-embedding-3-small** | 1536 | OpenAI apps, cost-effective |\n| **bge-large-en-v1.5** | 1024 | Open source, local deployment |\n| **multilingual-e5-large** | 1024 | Multi-language support |\n\n### 3. Retrieval Strategies\n\n**Approaches:**\n\n- **Dense Retrieval**: Semantic similarity via embeddings\n- **Sparse Retrieval**: Keyword matching (BM25, TF-IDF)\n- **Hybrid Search**: Combine dense + sparse with weighted fusion\n- **Multi-Query**: Generate multiple query variations\n- **HyDE**: Generate hypothetical documents for better retrieval\n\n### 4. Reranking\n\n**Purpose**: Improve retrieval quality by reordering results\n\n**Methods:**\n\n- **Cross-Encoders**: BERT-based reranking (ms-marco-MiniLM)\n- **Cohere Rerank**: API-based reranking\n- **Maximal Marginal Relevance (MMR)**: Diversity + relevance\n- **LLM-based**: Use LLM to score relevance\n\n## Quick Start with LangGraph\n\n```python\nfrom langgraph.graph import StateGraph, START, END\nfrom langchain_anthropic import ChatAnthropic\nfrom langchain_voyageai import VoyageAIEmbeddings\nfrom langchain_pinecone import PineconeVectorStore\nfrom langchain_core.documents import Document\nfrom langchain_core.prompts import ChatPromptTemplate\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\nfrom typing import TypedDict, Annotated\n\nclass RAGState(TypedDict):\n    question: str\n    context: list[Document]\n    answer: str\n\n# Initialize components\nllm = ChatAnthropic(model=\"claude-sonnet-4-6\")\nembeddings = VoyageAIEmbeddings(model=\"voyage-3-large\")\nvectorstore = PineconeVectorStore(index_name=\"docs\", embedding=embeddings)\nretriever = vectorstore.as_retriever(search_kwargs={\"k\": 4})\n\n# RAG prompt\nrag_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Answer based on the context below. If you cannot answer, say so.\n\n    Context:\n    {context}\n\n    Question: {question}\n\n    Answer:\"\"\"\n)\n\nasync def retrieve(state: RAGState) -> RAGState:\n    \"\"\"Retrieve relevant documents.\"\"\"\n    docs = await retriever.ainvoke(state[\"question\"])\n    return {\"context\": docs}\n\nasync def generate(state: RAGState) -> RAGState:\n    \"\"\"Generate answer from context.\"\"\"\n    context_text = \"\\n\\n\".join(doc.page_content for doc in state[\"context\"])\n    messages = rag_prompt.format_messages(\n        context=context_text,\n        question=state[\"question\"]\n    )\n    response = await llm.ainvoke(messages)\n    return {\"answer\": response.content}\n\n# Build RAG graph\nbuilder = StateGraph(RAGState)\nbuilder.add_node(\"retrieve\", retrieve)\nbuilder.add_node(\"generate\", generate)\nbuilder.add_edge(START, \"retrieve\")\nbuilder.add_edge(\"retrieve\", \"generate\")\nbuilder.add_edge(\"generate\", END)\n\nrag_chain = builder.compile()\n\n# Use\nresult = await rag_chain.ainvoke({\"question\": \"What are the main features?\"})\nprint(result[\"answer\"])\n```\n\n## Advanced RAG Patterns\n\n### Pattern 1: Hybrid Search with RRF\n\n```python\nfrom langchain_community.retrievers import BM25Retriever\nfrom langchain.retrievers import EnsembleRetriever\n\n# Sparse retriever (BM25 for keyword matching)\nbm25_retriever = BM25Retriever.from_documents(documents)\nbm25_retriever.k = 10\n\n# Dense retriever (embeddings for semantic search)\ndense_retriever = vectorstore.as_retriever(search_kwargs={\"k\": 10})\n\n# Combine with Reciprocal Rank Fusion weights\nensemble_retriever = EnsembleRetriever(\n    retrievers=[bm25_retriever, dense_retriever],\n    weights=[0.3, 0.7]  # 30% keyword, 70% semantic\n)\n```\n\n### Pattern 2: Multi-Query Retrieval\n\n```python\nfrom langchain.retrievers.multi_query import MultiQueryRetriever\n\n# Generate multiple query perspectives for better recall\nmulti_query_retriever = MultiQueryRetriever.from_llm(\n    retriever=vectorstore.as_retriever(search_kwargs={\"k\": 5}),\n    llm=llm\n)\n\n# Single query → multiple variations → combined results\nresults = await multi_query_retriever.ainvoke(\"What is the main topic?\")\n```\n\n### Pattern 3: Contextual Compression\n\n```python\nfrom langchain.retrievers import ContextualCompressionRetriever\nfrom langchain.retrievers.document_compressors import LLMChainExtractor\n\n# Compressor extracts only relevant portions\ncompressor = LLMChainExtractor.from_llm(llm)\n\ncompression_retriever = ContextualCompressionRetriever(\n    base_compressor=compressor,\n    base_retriever=vectorstore.as_retriever(search_kwargs={\"k\": 10})\n)\n\n# Returns only relevant parts of documents\ncompressed_docs = await compression_retriever.ainvoke(\"specific query\")\n```\n\n### Pattern 4: Parent Document Retriever\n\n```python\nfrom langchain.retrievers import ParentDocumentRetriever\nfrom langchain.storage import InMemoryStore\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\n\n# Small chunks for precise retrieval, large chunks for context\nchild_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=50)\nparent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)\n\n# Store for parent documents\ndocstore = InMemoryStore()\n\nparent_retriever = ParentDocumentRetriever(\n    vectorstore=vectorstore,\n    docstore=docstore,\n    child_splitter=child_splitter,\n    parent_splitter=parent_splitter\n)\n\n# Add documents (splits children, stores parents)\nawait parent_retriever.aadd_documents(documents)\n\n# Retrieval returns parent documents with full context\nresults = await parent_retriever.ainvoke(\"query\")\n```\n\n### Pattern 5: HyDE (Hypothetical Document Embeddings)\n\n```python\nfrom langchain_core.prompts import ChatPromptTemplate\n\nclass HyDEState(TypedDict):\n    question: str\n    hypothetical_doc: str\n    context: list[Document]\n    answer: str\n\nhyde_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Write a detailed passage that would answer this question:\n\n    Question: {question}\n\n    Passage:\"\"\"\n)\n\nasync def generate_hypothetical(state: HyDEState) -> HyDEState:\n    \"\"\"Generate hypothetical document for better retrieval.\"\"\"\n    messages = hyde_prompt.format_messages(question=state[\"question\"])\n    response = await llm.ainvoke(messages)\n    return {\"hypothetical_doc\": response.content}\n\nasync def retrieve_with_hyde(state: HyDEState) -> HyDEState:\n    \"\"\"Retrieve using hypothetical document.\"\"\"\n    # Use hypothetical doc for retrieval instead of original query\n    docs = await retriever.ainvoke(state[\"hypothetical_doc\"])\n    return {\"context\": docs}\n\n# Build HyDE RAG graph\nbuilder = StateGraph(HyDEState)\nbuilder.add_node(\"hypothetical\", generate_hypothetical)\nbuilder.add_node(\"retrieve\", retrieve_with_hyde)\nbuilder.add_node(\"generate\", generate)\nbuilder.add_edge(START, \"hypothetical\")\nbuilder.add_edge(\"hypothetical\", \"retrieve\")\nbuilder.add_edge(\"retrieve\", \"generate\")\nbuilder.add_edge(\"generate\", END)\n\nhyde_rag = builder.compile()\n```\n\n## Document Chunking Strategies\n\n### Recursive Character Text Splitter\n\n```python\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\n\nsplitter = RecursiveCharacterTextSplitter(\n    chunk_size=1000,\n    chunk_overlap=200,\n    length_function=len,\n    separators=[\"\\n\\n\", \"\\n\", \". \", \" \", \"\"]  # Try in order\n)\n\nchunks = splitter.split_documents(documents)\n```\n\n### Token-Based Splitting\n\n```python\nfrom langchain_text_splitters import TokenTextSplitter\n\nsplitter = TokenTextSplitter(\n    chunk_size=512,\n    chunk_overlap=50,\n    encoding_name=\"cl100k_base\"  # OpenAI tiktoken encoding\n)\n```\n\n### Semantic Chunking\n\n```python\nfrom langchain_experimental.text_splitter import SemanticChunker\n\nsplitter = SemanticChunker(\n    embeddings=embeddings,\n    breakpoint_threshold_type=\"percentile\",\n    breakpoint_threshold_amount=95\n)\n```\n\n### Markdown Header Splitter\n\n```python\nfrom langchain_text_splitters import MarkdownHeaderTextSplitter\n\nheaders_to_split_on = [\n    (\"#\", \"Header 1\"),\n    (\"##\", \"Header 2\"),\n    (\"###\", \"Header 3\"),\n]\n\nsplitter = MarkdownHeaderTextSplitter(\n    headers_to_split_on=headers_to_split_on,\n    strip_headers=False\n)\n```\n\n## Vector Store Configurations\n\n### Pinecone (Serverless)\n\n```python\nfrom pinecone import Pinecone, ServerlessSpec\nfrom langchain_pinecone import PineconeVectorStore\n\n# Initialize Pinecone client\npc = Pinecone(api_key=os.environ[\"PINECONE_API_KEY\"])\n\n# Create index if needed\nif \"my-index\" not in pc.list_indexes().names():\n    pc.create_index(\n        name=\"my-index\",\n        dimension=1024,  # voyage-3-large dimensions\n        metric=\"cosine\",\n        spec=ServerlessSpec(cloud=\"aws\", region=\"us-east-1\")\n    )\n\n# Create vector store\nindex = pc.Index(\"my-index\")\nvectorstore = PineconeVectorStore(index=index, embedding=embeddings)\n```\n\n### Weaviate\n\n```python\nimport weaviate\nfrom langchain_weaviate import WeaviateVectorStore\n\nclient = weaviate.connect_to_local()  # or connect_to_weaviate_cloud()\n\nvectorstore = WeaviateVectorStore(\n    client=client,\n    index_name=\"Documents\",\n    text_key=\"content\",\n    embedding=embeddings\n)\n```\n\n### Chroma (Local Development)\n\n```python\nfrom langchain_chroma import Chroma\n\nvectorstore = Chroma(\n    collection_name=\"my_collection\",\n    embedding_function=embeddings,\n    persist_directory=\".\u002Fchroma_db\"\n)\n```\n\n### pgvector (PostgreSQL)\n\n```python\nfrom langchain_postgres.vectorstores import PGVector\n\nconnection_string = \"postgresql+psycopg:\u002F\u002Fuser:pass@localhost:5432\u002Fvectordb\"\n\nvectorstore = PGVector(\n    embeddings=embeddings,\n    collection_name=\"documents\",\n    connection=connection_string,\n)\n```\n\n## Retrieval Optimization\n\n### 1. Metadata Filtering\n\n```python\nfrom langchain_core.documents import Document\n\n# Add metadata during indexing\ndocs_with_metadata = []\nfor doc in documents:\n    doc.metadata.update({\n        \"source\": doc.metadata.get(\"source\", \"unknown\"),\n        \"category\": determine_category(doc.page_content),\n        \"date\": datetime.now().isoformat()\n    })\n    docs_with_metadata.append(doc)\n\n# Filter during retrieval\nresults = await vectorstore.asimilarity_search(\n    \"query\",\n    filter={\"category\": \"technical\"},\n    k=5\n)\n```\n\n### 2. Maximal Marginal Relevance (MMR)\n\n```python\n# Balance relevance with diversity\nresults = await vectorstore.amax_marginal_relevance_search(\n    \"query\",\n    k=5,\n    fetch_k=20,  # Fetch 20, return top 5 diverse\n    lambda_mult=0.5  # 0=max diversity, 1=max relevance\n)\n```\n\n### 3. Reranking with Cross-Encoder\n\n```python\nfrom sentence_transformers import CrossEncoder\n\nreranker = CrossEncoder('cross-encoder\u002Fms-marco-MiniLM-L-6-v2')\n\nasync def retrieve_and_rerank(query: str, k: int = 5) -> list[Document]:\n    # Get initial results\n    candidates = await vectorstore.asimilarity_search(query, k=20)\n\n    # Rerank\n    pairs = [[query, doc.page_content] for doc in candidates]\n    scores = reranker.predict(pairs)\n\n    # Sort by score and take top k\n    ranked = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)\n    return [doc for doc, score in ranked[:k]]\n```\n\n### 4. Cohere Rerank\n\n```python\nfrom langchain.retrievers import CohereRerank\nfrom langchain_cohere import CohereRerank\n\nreranker = CohereRerank(model=\"rerank-english-v3.0\", top_n=5)\n\n# Wrap retriever with reranking\nreranked_retriever = ContextualCompressionRetriever(\n    base_compressor=reranker,\n    base_retriever=vectorstore.as_retriever(search_kwargs={\"k\": 20})\n)\n```\n\n## Prompt Engineering for RAG\n\n### Contextual Prompt with Citations\n\n```python\nrag_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Answer the question based on the context below. Include citations using [1], [2], etc.\n\n    If you cannot answer based on the context, say \"I don't have enough information.\"\n\n    Context:\n    {context}\n\n    Question: {question}\n\n    Instructions:\n    1. Use only information from the context\n    2. Cite sources with [1], [2] format\n    3. If uncertain, express uncertainty\n\n    Answer (with citations):\"\"\"\n)\n```\n\n### Structured Output for RAG\n\n```python\nfrom pydantic import BaseModel, Field\n\nclass RAGResponse(BaseModel):\n    answer: str = Field(description=\"The answer based on context\")\n    confidence: float = Field(description=\"Confidence score 0-1\")\n    sources: list[str] = Field(description=\"Source document IDs used\")\n    reasoning: str = Field(description=\"Brief reasoning for the answer\")\n\n# Use with structured output\nstructured_llm = llm.with_structured_output(RAGResponse)\n```\n\n## Evaluation Metrics\n\n```python\nfrom typing import TypedDict\n\nclass RAGEvalMetrics(TypedDict):\n    retrieval_precision: float  # Relevant docs \u002F retrieved docs\n    retrieval_recall: float     # Retrieved relevant \u002F total relevant\n    answer_relevance: float     # Answer addresses question\n    faithfulness: float         # Answer grounded in context\n    context_relevance: float    # Context relevant to question\n\nasync def evaluate_rag_system(\n    rag_chain,\n    test_cases: list[dict]\n) -> RAGEvalMetrics:\n    \"\"\"Evaluate RAG system on test cases.\"\"\"\n    metrics = {k: [] for k in RAGEvalMetrics.__annotations__}\n\n    for test in test_cases:\n        result = await rag_chain.ainvoke({\"question\": test[\"question\"]})\n\n        # Retrieval metrics\n        retrieved_ids = {doc.metadata[\"id\"] for doc in result[\"context\"]}\n        relevant_ids = set(test[\"relevant_doc_ids\"])\n\n        precision = len(retrieved_ids & relevant_ids) \u002F len(retrieved_ids)\n        recall = len(retrieved_ids & relevant_ids) \u002F len(relevant_ids)\n\n        metrics[\"retrieval_precision\"].append(precision)\n        metrics[\"retrieval_recall\"].append(recall)\n\n        # Use LLM-as-judge for quality metrics\n        quality = await evaluate_answer_quality(\n            question=test[\"question\"],\n            answer=result[\"answer\"],\n            context=result[\"context\"],\n            expected=test.get(\"expected_answer\")\n        )\n        metrics[\"answer_relevance\"].append(quality[\"relevance\"])\n        metrics[\"faithfulness\"].append(quality[\"faithfulness\"])\n        metrics[\"context_relevance\"].append(quality[\"context_relevance\"])\n\n    return {k: sum(v) \u002F len(v) for k, v in metrics.items()}\n```\n","# RAG Implementation\n\nMaster Retrieval-Augmented Generation (RAG) to build LLM applications that provide accurate, grounded responses using external knowledge sources.\n\n## When to Use This Skill\n\n- Building Q&A systems over proprietary documents\n- Creating chatbots with current, factual information\n- Implementing semantic search with natural language queries\n- Reducing hallucinations with grounded responses\n- Enabling LLMs to access domain-specific knowledge\n- Building documentation assistants\n- Creating research tools with source citation\n\n## Core Components\n\n### 1. Vector Databases\n\n**Purpose**: Store and retrieve document embeddings efficiently\n\n**Options:**\n\n- **Pinecone**: Managed, scalable, serverless\n- **Weaviate**: Open-source, hybrid search, GraphQL\n- **Milvus**: High performance, on-premise\n- **Chroma**: Lightweight, easy to use, local development\n- **Qdrant**: Fast, filtered search, Rust-based\n- **pgvector**: PostgreSQL extension, SQL integration\n\n### 2. Embeddings\n\n**Purpose**: Convert text to numerical vectors for similarity search\n\n**Models (2026):**\n| Model | Dimensions | Best For |\n|-------|------------|----------|\n| **voyage-3-large** | 1024 | Claude apps (Anthropic recommended) |\n| **voyage-code-3** | 1024 | Code search |\n| **text-embedding-3-large** | 3072 | OpenAI apps, high accuracy |\n| **text-embedding-3-small** | 1536 | OpenAI apps, cost-effective |\n| **bge-large-en-v1.5** | 1024 | Open source, local deployment |\n| **multilingual-e5-large** | 1024 | Multi-language support |\n\n### 3. Retrieval Strategies\n\n**Approaches:**\n\n- **Dense Retrieval**: Semantic similarity via embeddings\n- **Sparse Retrieval**: Keyword matching (BM25, TF-IDF)\n- **Hybrid Search**: Combine dense + sparse with weighted fusion\n- **Multi-Query**: Generate multiple query variations\n- **HyDE**: Generate hypothetical documents for better retrieval\n\n### 4. Reranking\n\n**Purpose**: Improve retrieval quality by reordering results\n\n**Methods:**\n\n- **Cross-Encoders**: BERT-based reranking (ms-marco-MiniLM)\n- **Cohere Rerank**: API-based reranking\n- **Maximal Marginal Relevance (MMR)**: Diversity + relevance\n- **LLM-based**: Use LLM to score relevance\n\n## Quick Start with LangGraph\n\n```python\nfrom langgraph.graph import StateGraph, START, END\nfrom langchain_anthropic import ChatAnthropic\nfrom langchain_voyageai import VoyageAIEmbeddings\nfrom langchain_pinecone import PineconeVectorStore\nfrom langchain_core.documents import Document\nfrom langchain_core.prompts import ChatPromptTemplate\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\nfrom typing import TypedDict, Annotated\n\nclass RAGState(TypedDict):\n    question: str\n    context: list[Document]\n    answer: str\n\n# Initialize components\nllm = ChatAnthropic(model=\"claude-sonnet-4-6\")\nembeddings = VoyageAIEmbeddings(model=\"voyage-3-large\")\nvectorstore = PineconeVectorStore(index_name=\"docs\", embedding=embeddings)\nretriever = vectorstore.as_retriever(search_kwargs={\"k\": 4})\n\n# RAG prompt\nrag_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Answer based on the context below. If you cannot answer, say so.\n\n    Context:\n    {context}\n\n    Question: {question}\n\n    Answer:\"\"\"\n)\n\nasync def retrieve(state: RAGState) -> RAGState:\n    \"\"\"Retrieve relevant documents.\"\"\"\n    docs = await retriever.ainvoke(state[\"question\"])\n    return {\"context\": docs}\n\nasync def generate(state: RAGState) -> RAGState:\n    \"\"\"Generate answer from context.\"\"\"\n    context_text = \"\\n\\n\".join(doc.page_content for doc in state[\"context\"])\n    messages = rag_prompt.format_messages(\n        context=context_text,\n        question=state[\"question\"]\n    )\n    response = await llm.ainvoke(messages)\n    return {\"answer\": response.content}\n\n# Build RAG graph\nbuilder = StateGraph(RAGState)\nbuilder.add_node(\"retrieve\", retrieve)\nbuilder.add_node(\"generate\", generate)\nbuilder.add_edge(START, \"retrieve\")\nbuilder.add_edge(\"retrieve\", \"generate\")\nbuilder.add_edge(\"generate\", END)\n\nrag_chain = builder.compile()\n\n# Use\nresult = await rag_chain.ainvoke({\"question\": \"What are the main features?\"})\nprint(result[\"answer\"])\n```\n\n## Advanced RAG Patterns\n\n### Pattern 1: Hybrid Search with RRF\n\n```python\nfrom langchain_community.retrievers import BM25Retriever\nfrom langchain.retrievers import EnsembleRetriever\n\n# Sparse retriever (BM25 for keyword matching)\nbm25_retriever = BM25Retriever.from_documents(documents)\nbm25_retriever.k = 10\n\n# Dense retriever (embeddings for semantic search)\ndense_retriever = vectorstore.as_retriever(search_kwargs={\"k\": 10})\n\n# Combine with Reciprocal Rank Fusion weights\nensemble_retriever = EnsembleRetriever(\n    retrievers=[bm25_retriever, dense_retriever],\n    weights=[0.3, 0.7]  # 30% keyword, 70% semantic\n)\n```\n\n### Pattern 2: Multi-Query Retrieval\n\n```python\nfrom langchain.retrievers.multi_query import MultiQueryRetriever\n\n# Generate multiple query perspectives for better recall\nmulti_query_retriever = MultiQueryRetriever.from_llm(\n    retriever=vectorstore.as_retriever(search_kwargs={\"k\": 5}),\n    llm=llm\n)\n\n# Single query → multiple variations → combined results\nresults = await multi_query_retriever.ainvoke(\"What is the main topic?\")\n```\n\n### Pattern 3: Contextual Compression\n\n```python\nfrom langchain.retrievers import ContextualCompressionRetriever\nfrom langchain.retrievers.document_compressors import LLMChainExtractor\n\n# Compressor extracts only relevant portions\ncompressor = LLMChainExtractor.from_llm(llm)\n\ncompression_retriever = ContextualCompressionRetriever(\n    base_compressor=compressor,\n    base_retriever=vectorstore.as_retriever(search_kwargs={\"k\": 10})\n)\n\n# Returns only relevant parts of documents\ncompressed_docs = await compression_retriever.ainvoke(\"specific query\")\n```\n\n### Pattern 4: Parent Document Retriever\n\n```python\nfrom langchain.retrievers import ParentDocumentRetriever\nfrom langchain.storage import InMemoryStore\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\n\n# Small chunks for precise retrieval, large chunks for context\nchild_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=50)\nparent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)\n\n# Store for parent documents\ndocstore = InMemoryStore()\n\nparent_retriever = ParentDocumentRetriever(\n    vectorstore=vectorstore,\n    docstore=docstore,\n    child_splitter=child_splitter,\n    parent_splitter=parent_splitter\n)\n\n# Add documents (splits children, stores parents)\nawait parent_retriever.aadd_documents(documents)\n\n# Retrieval returns parent documents with full context\nresults = await parent_retriever.ainvoke(\"query\")\n```\n\n### Pattern 5: HyDE (Hypothetical Document Embeddings)\n\n```python\nfrom langchain_core.prompts import ChatPromptTemplate\n\nclass HyDEState(TypedDict):\n    question: str\n    hypothetical_doc: str\n    context: list[Document]\n    answer: str\n\nhyde_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Write a detailed passage that would answer this question:\n\n    Question: {question}\n\n    Passage:\"\"\"\n)\n\nasync def generate_hypothetical(state: HyDEState) -> HyDEState:\n    \"\"\"Generate hypothetical document for better retrieval.\"\"\"\n    messages = hyde_prompt.format_messages(question=state[\"question\"])\n    response = await llm.ainvoke(messages)\n    return {\"hypothetical_doc\": response.content}\n\nasync def retrieve_with_hyde(state: HyDEState) -> HyDEState:\n    \"\"\"Retrieve using hypothetical document.\"\"\"\n    # Use hypothetical doc for retrieval instead of original query\n    docs = await retriever.ainvoke(state[\"hypothetical_doc\"])\n    return {\"context\": docs}\n\n# Build HyDE RAG graph\nbuilder = StateGraph(HyDEState)\nbuilder.add_node(\"hypothetical\", generate_hypothetical)\nbuilder.add_node(\"retrieve\", retrieve_with_hyde)\nbuilder.add_node(\"generate\", generate)\nbuilder.add_edge(START, \"hypothetical\")\nbuilder.add_edge(\"hypothetical\", \"retrieve\")\nbuilder.add_edge(\"retrieve\", \"generate\")\nbuilder.add_edge(\"generate\", END)\n\nhyde_rag = builder.compile()\n```\n\n## Document Chunking Strategies\n\n### Recursive Character Text Splitter\n\n```python\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\n\nsplitter = RecursiveCharacterTextSplitter(\n    chunk_size=1000,\n    chunk_overlap=200,\n    length_function=len,\n    separators=[\"\\n\\n\", \"\\n\", \". \", \" \", \"\"]  # Try in order\n)\n\nchunks = splitter.split_documents(documents)\n```\n\n### Token-Based Splitting\n\n```python\nfrom langchain_text_splitters import TokenTextSplitter\n\nsplitter = TokenTextSplitter(\n    chunk_size=512,\n    chunk_overlap=50,\n    encoding_name=\"cl100k_base\"  # OpenAI tiktoken encoding\n)\n```\n\n### Semantic Chunking\n\n```python\nfrom langchain_experimental.text_splitter import SemanticChunker\n\nsplitter = SemanticChunker(\n    embeddings=embeddings,\n    breakpoint_threshold_type=\"percentile\",\n    breakpoint_threshold_amount=95\n)\n```\n\n### Markdown Header Splitter\n\n```python\nfrom langchain_text_splitters import MarkdownHeaderTextSplitter\n\nheaders_to_split_on = [\n    (\"#\", \"Header 1\"),\n    (\"##\", \"Header 2\"),\n    (\"###\", \"Header 3\"),\n]\n\nsplitter = MarkdownHeaderTextSplitter(\n    headers_to_split_on=headers_to_split_on,\n    strip_headers=False\n)\n```\n\n## Vector Store Configurations\n\n### Pinecone (Serverless)\n\n```python\nfrom pinecone import Pinecone, ServerlessSpec\nfrom langchain_pinecone import PineconeVectorStore\n\n# Initialize Pinecone client\npc = Pinecone(api_key=os.environ[\"PINECONE_API_KEY\"])\n\n# Create index if needed\nif \"my-index\" not in pc.list_indexes().names():\n    pc.create_index(\n        name=\"my-index\",\n        dimension=1024,  # voyage-3-large dimensions\n        metric=\"cosine\",\n        spec=ServerlessSpec(cloud=\"aws\", region=\"us-east-1\")\n    )\n\n# Create vector store\nindex = pc.Index(\"my-index\")\nvectorstore = PineconeVectorStore(index=index, embedding=embeddings)\n```\n\n### Weaviate\n\n```python\nimport weaviate\nfrom langchain_weaviate import WeaviateVectorStore\n\nclient = weaviate.connect_to_local()  # or connect_to_weaviate_cloud()\n\nvectorstore = WeaviateVectorStore(\n    client=client,\n    index_name=\"Documents\",\n    text_key=\"content\",\n    embedding=embeddings\n)\n```\n\n### Chroma (Local Development)\n\n```python\nfrom langchain_chroma import Chroma\n\nvectorstore = Chroma(\n    collection_name=\"my_collection\",\n    embedding_function=embeddings,\n    persist_directory=\".\u002Fchroma_db\"\n)\n```\n\n### pgvector (PostgreSQL)\n\n```python\nfrom langchain_postgres.vectorstores import PGVector\n\nconnection_string = \"postgresql+psycopg:\u002F\u002Fuser:pass@localhost:5432\u002Fvectordb\"\n\nvectorstore = PGVector(\n    embeddings=embeddings,\n    collection_name=\"documents\",\n    connection=connection_string,\n)\n```\n\n## Retrieval Optimization\n\n### 1. Metadata Filtering\n\n```python\nfrom langchain_core.documents import Document\n\n# Add metadata during indexing\ndocs_with_metadata = []\nfor doc in documents:\n    doc.metadata.update({\n        \"source\": doc.metadata.get(\"source\", \"unknown\"),\n        \"category\": determine_category(doc.page_content),\n        \"date\": datetime.now().isoformat()\n    })\n    docs_with_metadata.append(doc)\n\n# Filter during retrieval\nresults = await vectorstore.asimilarity_search(\n    \"query\",\n    filter={\"category\": \"technical\"},\n    k=5\n)\n```\n\n### 2. Maximal Marginal Relevance (MMR)\n\n```python\n# Balance relevance with diversity\nresults = await vectorstore.amax_marginal_relevance_search(\n    \"query\",\n    k=5,\n    fetch_k=20,  # Fetch 20, return top 5 diverse\n    lambda_mult=0.5  # 0=max diversity, 1=max relevance\n)\n```\n\n### 3. Reranking with Cross-Encoder\n\n```python\nfrom sentence_transformers import CrossEncoder\n\nreranker = CrossEncoder('cross-encoder\u002Fms-marco-MiniLM-L-6-v2')\n\nasync def retrieve_and_rerank(query: str, k: int = 5) -> list[Document]:\n    # Get initial results\n    candidates = await vectorstore.asimilarity_search(query, k=20)\n\n    # Rerank\n    pairs = [[query, doc.page_content] for doc in candidates]\n    scores = reranker.predict(pairs)\n\n    # Sort by score and take top k\n    ranked = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)\n    return [doc for doc, score in ranked[:k]]\n```\n\n### 4. Cohere Rerank\n\n```python\nfrom langchain.retrievers import CohereRerank\nfrom langchain_cohere import CohereRerank\n\nreranker = CohereRerank(model=\"rerank-english-v3.0\", top_n=5)\n\n# Wrap retriever with reranking\nreranked_retriever = ContextualCompressionRetriever(\n    base_compressor=reranker,\n    base_retriever=vectorstore.as_retriever(search_kwargs={\"k\": 20})\n)\n```\n\n## Prompt Engineering for RAG\n\n### Contextual Prompt with Citations\n\n```python\nrag_prompt = ChatPromptTemplate.from_template(\n    \"\"\"Answer the question based on the context below. Include citations using [1], [2], etc.\n\n    If you cannot answer based on the context, say \"I don't have enough information.\"\n\n    Context:\n    {context}\n\n    Question: {question}\n\n    Instructions:\n    1. Use only information from the context\n    2. Cite sources with [1], [2] format\n    3. If uncertain, express uncertainty\n\n    Answer (with citations):\"\"\"\n)\n```\n\n### Structured Output for RAG\n\n```python\nfrom pydantic import BaseModel, Field\n\nclass RAGResponse(BaseModel):\n    answer: str = Field(description=\"The answer based on context\")\n    confidence: float = Field(description=\"Confidence score 0-1\")\n    sources: list[str] = Field(description=\"Source document IDs used\")\n    reasoning: str = Field(description=\"Brief reasoning for the answer\")\n\n# Use with structured output\nstructured_llm = llm.with_structured_output(RAGResponse)\n```\n\n## Evaluation Metrics\n\n```python\nfrom typing import TypedDict\n\nclass RAGEvalMetrics(TypedDict):\n    retrieval_precision: float  # Relevant docs \u002F retrieved docs\n    retrieval_recall: float     # Retrieved relevant \u002F total relevant\n    answer_relevance: float     # Answer addresses question\n    faithfulness: float         # Answer grounded in context\n    context_relevance: float    # Context relevant to question\n\nasync def evaluate_rag_system(\n    rag_chain,\n    test_cases: list[dict]\n) -> RAGEvalMetrics:\n    \"\"\"Evaluate RAG system on test cases.\"\"\"\n    metrics = {k: [] for k in RAGEvalMetrics.__annotations__}\n\n    for test in test_cases:\n        result = await rag_chain.ainvoke({\"question\": test[\"question\"]})\n\n        # Retrieval metrics\n        retrieved_ids = {doc.metadata[\"id\"] for doc in result[\"context\"]}\n        relevant_ids = set(test[\"relevant_doc_ids\"])\n\n        precision = len(retrieved_ids & relevant_ids) \u002F len(retrieved_ids)\n        recall = len(retrieved_ids & relevant_ids) \u002F len(relevant_ids)\n\n        metrics[\"retrieval_precision\"].append(precision)\n        metrics[\"retrieval_recall\"].append(recall)\n\n        # Use LLM-as-judge for quality metrics\n        quality = await evaluate_answer_quality(\n            question=test[\"question\"],\n            answer=result[\"answer\"],\n            context=result[\"context\"],\n            expected=test.get(\"expected_answer\")\n        )\n        metrics[\"answer_relevance\"].append(quality[\"relevance\"])\n        metrics[\"faithfulness\"].append(quality[\"faithfulness\"])\n        metrics[\"context_relevance\"].append(quality[\"context_relevance\"])\n\n    return {k: sum(v) \u002F len(v) for k, v in metrics.items()}\n```\n",{"name":5,"description":7},"main","900bb95bbe72cdb94af6f5a620aa5081264b7b6d",".claude\u002Fmarketplaces\u002Fwshobson\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Frag-implementation",0,"https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up#readme","https:\u002F\u002Fgithub.com\u002Fsteph-frtech","2026-05-16 00:03:30",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add steph-frtech\u002Fclaude-code-up --skill rag-implementation",{"tool":37,"command":38},"bunx","bunx skills add steph-frtech\u002Fclaude-code-up --skill rag-implementation",{"tool":40,"command":41},"pnpm","pnpm dlx skills add steph-frtech\u002Fclaude-code-up --skill rag-implementation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fmarketplaces\u002Fwshobson\u002Fplugins\u002Fllm-application-dev\u002Fskills\u002Frag-implementation\u002FSKILL.md",true,[],[46,53,60,67,74,81],{"skillNo":47,"name":48,"nameZh":6,"description":49,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":50,"categories":51,"keywords":52,"publishedAt":18},"01M72Y1BY6X8TX3DS19Y04NEBQ","example","One-line summary of what this skill does","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":18},"01M72Y1A38EYR8S9P1MBYDAQTW","writing-plans","Use when you have a spec or requirements for a multi-step task, before touching code","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsuperpowers__writing-plans\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":18},"01M72Y174QSV1B5NV1E4KWHQBR","using-git-worktrees","Use when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsuperpowers__using-git-worktrees\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":18},"01M72Y14SQRHHMKG316HGB2CGK","test-driven-development","Use when implementing any feature or bugfix, before writing implementation code","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsuperpowers__test-driven-development\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":18},"01M72Y127BCPRH3BRKPJC3Q3G2","systematic-debugging","Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsuperpowers__systematic-debugging\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":18},"01M72Y1090GJA6B67KPD60WPHH","subagent-driven-development","Use when executing implementation plans with independent tasks in the current session","https:\u002F\u002Fgithub.com\u002Fsteph-frtech\u002Fclaude-code-up\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsuperpowers__subagent-driven-development\u002FSKILL.md",[],[]]