[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6NW3VT7PD7BR2N36T384GRG":3,"skill-same:01M6NW3VT7PD7BR2N36T384GRG":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":33,"linkOnly":34,"markdownContent":35,"markdownBody":36,"frontMatter":37,"refName":38,"commitSha":39,"skillDirectory":40,"forksCount":14,"repositoryHtmlUrl":41,"repositoryHomepageUrl":6,"ownerHtmlUrl":42,"ownerBlogUrl":43,"ownerDescription":44,"updatedAt":45,"install":46,"resources":59},"01M6NW3VT7PD7BR2N36T384GRG","rag-patterns",null,"Chunking strategies, embedding model selection, hybrid search, reranking, eval metrics","rubicanjr\u002FFinCognis","rubicanjr","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F234912519?v=4","01M6NT156NND8J474G34SB3923","Rubi","01M6NT15B70RWXE1N3YV0BZY8A",0,"MIT","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Frag-patterns\u002Fskill.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32],"rag","patterns","rag patterns","model selection","metrics","search","model","eval","selection","strategies","chunking","embedding","Hybrid Search","hybrid","2026-11-10 10:12:30",false,"---\nname: rag-patterns\ndescription: Chunking strategies, embedding model selection, hybrid search, reranking, eval metrics\n---\n\n# RAG Patterns\n\n## Chunking Strategies\n\n```python\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\n\n# Strategy 1: Recursive character splitting (general purpose)\nsplitter = RecursiveCharacterTextSplitter(\n    chunk_size=512,\n    chunk_overlap=64,\n    separators=[\"\\n\\n\", \"\\n\", \". \", \" \", \"\"],\n    length_function=len,\n)\n\n# Strategy 2: Semantic chunking (better coherence)\nfrom langchain_experimental.text_splitter import SemanticChunker\nfrom langchain_openai import OpenAIEmbeddings\n\nsemantic_splitter = SemanticChunker(\n    OpenAIEmbeddings(),\n    breakpoint_threshold_type=\"percentile\",\n    breakpoint_threshold_amount=95,\n)\n\n# Strategy 3: Parent-child chunking (preserves context)\nparent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)\nchild_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=50)\n\nparent_docs = parent_splitter.split_documents(documents)\nfor parent in parent_docs:\n    children = child_splitter.split_documents([parent])\n    for child in children:\n        child.metadata[\"parent_id\"] = parent.metadata[\"id\"]\n```\n\n## Embedding Model Selection\n\n```yaml\nModels by Use Case:\n  General (English):\n    - text-embedding-3-small (OpenAI, 1536d, cheap)\n    - text-embedding-3-large (OpenAI, 3072d, best quality)\n    - all-MiniLM-L6-v2 (local, 384d, fast)\n  Code:\n    - text-embedding-3-large with code-tuned prompts\n    - voyage-code-2 (Voyage AI)\n  Multilingual:\n    - multilingual-e5-large (local)\n    - text-embedding-3-large (OpenAI)\n\nSelection Criteria:\n  - Latency requirement \u003C 50ms → local model\n  - Quality critical → text-embedding-3-large\n  - Budget constrained → text-embedding-3-small\n  - Air-gapped → all-MiniLM-L6-v2\n```\n\n## Hybrid Search (Vector + BM25)\n\n```python\nfrom rank_bm25 import BM25Okapi\nimport numpy as np\n\nclass HybridRetriever:\n    def __init__(self, vector_store, documents, alpha=0.5):\n        self.vector_store = vector_store\n        self.alpha = alpha  # 0=BM25 only, 1=vector only\n        tokenized = [doc.page_content.lower().split() for doc in documents]\n        self.bm25 = BM25Okapi(tokenized)\n        self.documents = documents\n\n    def search(self, query: str, k: int = 10) -> list:\n        # Vector search\n        vector_results = self.vector_store.similarity_search_with_score(query, k=k)\n        vector_scores = {doc.metadata[\"id\"]: score for doc, score in vector_results}\n\n        # BM25 search\n        bm25_scores_raw = self.bm25.get_scores(query.lower().split())\n        bm25_max = max(bm25_scores_raw) if max(bm25_scores_raw) > 0 else 1\n        bm25_scores = {\n            self.documents[i].metadata[\"id\"]: score \u002F bm25_max\n            for i, score in enumerate(bm25_scores_raw)\n        }\n\n        # Reciprocal Rank Fusion\n        all_ids = set(vector_scores) | set(bm25_scores)\n        fused = {}\n        for doc_id in all_ids:\n            v_score = vector_scores.get(doc_id, 0)\n            b_score = bm25_scores.get(doc_id, 0)\n            fused[doc_id] = self.alpha * v_score + (1 - self.alpha) * b_score\n\n        sorted_ids = sorted(fused, key=fused.get, reverse=True)[:k]\n        return [self._get_doc(did) for did in sorted_ids]\n```\n\n## Reranking\n\n```python\nfrom sentence_transformers import CrossEncoder\n\nreranker = CrossEncoder(\"cross-encoder\u002Fms-marco-MiniLM-L-6-v2\")\n\ndef rerank(query: str, documents: list, top_k: int = 5) -> list:\n    pairs = [(query, doc.page_content) for doc in documents]\n    scores = reranker.predict(pairs)\n    ranked = sorted(zip(documents, scores), key=lambda x: x[1], reverse=True)\n    return [doc for doc, _ in ranked[:top_k]]\n\n# Pipeline: retrieve 20 → rerank to 5\ncandidates = hybrid_retriever.search(query, k=20)\nfinal = rerank(query, candidates, top_k=5)\n```\n\n## Evaluation Metrics\n\n```python\ndef evaluate_rag(queries, expected_answers, retriever, generator):\n    metrics = {\"retrieval_recall\": [], \"answer_correctness\": [], \"faithfulness\": []}\n\n    for query, expected in zip(queries, expected_answers):\n        retrieved = retriever.search(query, k=5)\n        retrieved_texts = [d.page_content for d in retrieved]\n\n        # Retrieval recall: did we find the right chunks?\n        relevant_found = any(expected[\"source\"] in t for t in retrieved_texts)\n        metrics[\"retrieval_recall\"].append(1.0 if relevant_found else 0.0)\n\n        # Generate answer\n        answer = generator.generate(query, retrieved_texts)\n\n        # Faithfulness: is answer grounded in retrieved context?\n        # (Use LLM-as-judge or NLI model)\n        metrics[\"faithfulness\"].append(check_faithfulness(answer, retrieved_texts))\n\n        # Correctness: does answer match expected?\n        metrics[\"answer_correctness\"].append(check_correctness(answer, expected[\"answer\"]))\n\n    return {k: sum(v) \u002F len(v) for k, v in metrics.items()}\n```\n\n## Checklist\n\n- [ ] Chunk size tuned for domain (code: 1000+, prose: 300-500)\n- [ ] Chunk overlap prevents context loss at boundaries\n- [ ] Metadata preserved (source, page, section) for attribution\n- [ ] Hybrid search combines vector + keyword for robustness\n- [ ] Reranker applied before final context assembly\n- [ ] Context window budget managed (don't exceed LLM limit)\n- [ ] Evaluation pipeline with retrieval recall + faithfulness\n- [ ] Embedding model benchmarked on domain-specific queries\n\n## Anti-Patterns\n\n- Chunking without overlap (losing context at boundaries)\n- Using only vector search (misses exact keyword matches)\n- Stuffing all retrieved chunks into prompt (exceeds context window)\n- No evaluation pipeline (can't measure improvements)\n- Embedding queries and documents with different models\n- Ignoring metadata filtering before vector search\n- Not handling empty retrieval results gracefully\n","# RAG Patterns\n\n## Chunking Strategies\n\n```python\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\n\n# Strategy 1: Recursive character splitting (general purpose)\nsplitter = RecursiveCharacterTextSplitter(\n    chunk_size=512,\n    chunk_overlap=64,\n    separators=[\"\\n\\n\", \"\\n\", \". \", \" \", \"\"],\n    length_function=len,\n)\n\n# Strategy 2: Semantic chunking (better coherence)\nfrom langchain_experimental.text_splitter import SemanticChunker\nfrom langchain_openai import OpenAIEmbeddings\n\nsemantic_splitter = SemanticChunker(\n    OpenAIEmbeddings(),\n    breakpoint_threshold_type=\"percentile\",\n    breakpoint_threshold_amount=95,\n)\n\n# Strategy 3: Parent-child chunking (preserves context)\nparent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)\nchild_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=50)\n\nparent_docs = parent_splitter.split_documents(documents)\nfor parent in parent_docs:\n    children = child_splitter.split_documents([parent])\n    for child in children:\n        child.metadata[\"parent_id\"] = parent.metadata[\"id\"]\n```\n\n## Embedding Model Selection\n\n```yaml\nModels by Use Case:\n  General (English):\n    - text-embedding-3-small (OpenAI, 1536d, cheap)\n    - text-embedding-3-large (OpenAI, 3072d, best quality)\n    - all-MiniLM-L6-v2 (local, 384d, fast)\n  Code:\n    - text-embedding-3-large with code-tuned prompts\n    - voyage-code-2 (Voyage AI)\n  Multilingual:\n    - multilingual-e5-large (local)\n    - text-embedding-3-large (OpenAI)\n\nSelection Criteria:\n  - Latency requirement \u003C 50ms → local model\n  - Quality critical → text-embedding-3-large\n  - Budget constrained → text-embedding-3-small\n  - Air-gapped → all-MiniLM-L6-v2\n```\n\n## Hybrid Search (Vector + BM25)\n\n```python\nfrom rank_bm25 import BM25Okapi\nimport numpy as np\n\nclass HybridRetriever:\n    def __init__(self, vector_store, documents, alpha=0.5):\n        self.vector_store = vector_store\n        self.alpha = alpha  # 0=BM25 only, 1=vector only\n        tokenized = [doc.page_content.lower().split() for doc in documents]\n        self.bm25 = BM25Okapi(tokenized)\n        self.documents = documents\n\n    def search(self, query: str, k: int = 10) -> list:\n        # Vector search\n        vector_results = self.vector_store.similarity_search_with_score(query, k=k)\n        vector_scores = {doc.metadata[\"id\"]: score for doc, score in vector_results}\n\n        # BM25 search\n        bm25_scores_raw = self.bm25.get_scores(query.lower().split())\n        bm25_max = max(bm25_scores_raw) if max(bm25_scores_raw) > 0 else 1\n        bm25_scores = {\n            self.documents[i].metadata[\"id\"]: score \u002F bm25_max\n            for i, score in enumerate(bm25_scores_raw)\n        }\n\n        # Reciprocal Rank Fusion\n        all_ids = set(vector_scores) | set(bm25_scores)\n        fused = {}\n        for doc_id in all_ids:\n            v_score = vector_scores.get(doc_id, 0)\n            b_score = bm25_scores.get(doc_id, 0)\n            fused[doc_id] = self.alpha * v_score + (1 - self.alpha) * b_score\n\n        sorted_ids = sorted(fused, key=fused.get, reverse=True)[:k]\n        return [self._get_doc(did) for did in sorted_ids]\n```\n\n## Reranking\n\n```python\nfrom sentence_transformers import CrossEncoder\n\nreranker = CrossEncoder(\"cross-encoder\u002Fms-marco-MiniLM-L-6-v2\")\n\ndef rerank(query: str, documents: list, top_k: int = 5) -> list:\n    pairs = [(query, doc.page_content) for doc in documents]\n    scores = reranker.predict(pairs)\n    ranked = sorted(zip(documents, scores), key=lambda x: x[1], reverse=True)\n    return [doc for doc, _ in ranked[:top_k]]\n\n# Pipeline: retrieve 20 → rerank to 5\ncandidates = hybrid_retriever.search(query, k=20)\nfinal = rerank(query, candidates, top_k=5)\n```\n\n## Evaluation Metrics\n\n```python\ndef evaluate_rag(queries, expected_answers, retriever, generator):\n    metrics = {\"retrieval_recall\": [], \"answer_correctness\": [], \"faithfulness\": []}\n\n    for query, expected in zip(queries, expected_answers):\n        retrieved = retriever.search(query, k=5)\n        retrieved_texts = [d.page_content for d in retrieved]\n\n        # Retrieval recall: did we find the right chunks?\n        relevant_found = any(expected[\"source\"] in t for t in retrieved_texts)\n        metrics[\"retrieval_recall\"].append(1.0 if relevant_found else 0.0)\n\n        # Generate answer\n        answer = generator.generate(query, retrieved_texts)\n\n        # Faithfulness: is answer grounded in retrieved context?\n        # (Use LLM-as-judge or NLI model)\n        metrics[\"faithfulness\"].append(check_faithfulness(answer, retrieved_texts))\n\n        # Correctness: does answer match expected?\n        metrics[\"answer_correctness\"].append(check_correctness(answer, expected[\"answer\"]))\n\n    return {k: sum(v) \u002F len(v) for k, v in metrics.items()}\n```\n\n## Checklist\n\n- [ ] Chunk size tuned for domain (code: 1000+, prose: 300-500)\n- [ ] Chunk overlap prevents context loss at boundaries\n- [ ] Metadata preserved (source, page, section) for attribution\n- [ ] Hybrid search combines vector + keyword for robustness\n- [ ] Reranker applied before final context assembly\n- [ ] Context window budget managed (don't exceed LLM limit)\n- [ ] Evaluation pipeline with retrieval recall + faithfulness\n- [ ] Embedding model benchmarked on domain-specific queries\n\n## Anti-Patterns\n\n- Chunking without overlap (losing context at boundaries)\n- Using only vector search (misses exact keyword matches)\n- Stuffing all retrieved chunks into prompt (exceeds context window)\n- No evaluation pipeline (can't measure improvements)\n- Embedding queries and documents with different models\n- Ignoring metadata filtering before vector search\n- Not handling empty retrieval results gracefully\n",{"name":5,"description":7},"main","70e1c350500ee750987420bbae48837afa906b7e","skills\u002Frag-patterns","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis","https:\u002F\u002Fgithub.com\u002Frubicanjr","linktr.ee\u002Ffinanskodu","Industrial Engineer & CIO @FinansKodu. Architecting AI-driven financial agents and autonomous workflows. 🏗️ Building systems, not just prompts.","2026-06-09 13:50:15",{"source":8,"skillName":5,"commands":47,"prompt":57,"zipAvailable":58},[48,51,54],{"tool":49,"command":50},"npx","npx skills add rubicanjr\u002FFinCognis --skill rag-patterns",{"tool":52,"command":53},"bunx","bunx skills add rubicanjr\u002FFinCognis --skill rag-patterns",{"tool":55,"command":56},"pnpm","pnpm dlx skills add rubicanjr\u002FFinCognis --skill rag-patterns","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Frag-patterns\u002Fskill.md",true,[],[61,82,91,109,139,158],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":72,"publishedAt":33},"01M6NWS9DAXJTADT4T5MB49X8W","workflow-router","Goal-based workflow orchestration - routes tasks to specialist agents based on user goals","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fworkflow-router\u002FSKILL.md",[67],{"categoryNo":68,"categoryName":69,"taxonomyType":70,"taxonomyTypeDesc":71},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[73,74,75,76,77,78,79,80,81],"workflow router","workflow","router","orchestration","goals","goal","user","workflow orchestration","tasks",{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":33},"01M6NWS3RETSC0SHP16HP66E14","wiring","Wiring Verification","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fwiring\u002FSKILL.md",[],[84,89,90],"verification","false",{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":100,"publishedAt":33},"01M6NWS0XW9CVQZBDA1EGW0TWC","websocket-patterns","Connection management, room patterns, reconnection strategies, message buffering, and binary protocol design.","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fwebsocket-patterns\u002Fskill.md",[97],{"categoryNo":98,"categoryName":99,"taxonomyType":70,"taxonomyTypeDesc":71},"01JBSEEDCAT000000000000015","视觉设计",[101,20,102,103,104,105,106,28,107,108],"websocket","websocket patterns","design","protocol","management","message","binary","connection",{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":119,"publishedAt":33},"01M6NWRDQDX9ZNETAWRRMQWV52","visual-verdict","Screenshot comparison QA for frontend development. Takes a screenshot of the current implementation, scores it across multiple visual dimensions, and returns a structured PASS\u002FREVISE\u002FFAIL verdict with concrete fixes. Use when implementing UI from a design reference or verifying visual correctness.","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fvisual-verdict\u002FSKILL.md",[115,118],{"categoryNo":116,"categoryName":117,"taxonomyType":70,"taxonomyTypeDesc":71},"01JBSEEDCAT000000000000009","前端开发",{"categoryNo":68,"categoryName":69,"taxonomyType":70,"taxonomyTypeDesc":71},[120,121,122,123,124,125,126,127,128,129,130,131,132,103,133,134,135,136,137,138],"visual","verdict","visual verdict","ui","across","revise","comparison","correctness","implementing","for","fixes","dimensions","frontend","development","qa","screenshot","returns","use","frontend development",{"skillNo":140,"name":141,"nameZh":6,"description":142,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":143,"categories":144,"keywords":145,"publishedAt":33},"01M6NWR8HSNRS7MNDK2J02QYMN","verification-loop","Comprehensive verification system covering build, types, lint, tests, security, and diff review before a PR.","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fverification-loop\u002FSKILL.md",[],[146,147,89,148,149,150,151,152,153,154,155,156,157],"verification loop","loop","lint","review","security","pr","system","diff","build","tests","diff review","types",{"skillNo":159,"name":160,"nameZh":6,"description":161,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":162,"categories":163,"keywords":164,"publishedAt":33},"01M6NWR4HHSFR0VR5EXFRWA747","vector-db-patterns","Embedding strategies, ANN algorithms, hybrid search, RAG chunking strategies, and reranking for semantic search and retrieval.","https:\u002F\u002Fgithub.com\u002Frubicanjr\u002FFinCognis\u002Fblob\u002Fmain\u002Fskills\u002Fvector-db-patterns\u002Fskill.md",[],[165,166,167,168,20,169,24,170,171,28,29,172,173,30,31,129,32,19],"vector db patterns","vector db","db","vector","semantic search","embedding strategies","algorithms","semantic","retrieval"]