rag-implementation

Name: rag-implementation
Author: aisa-group

88stars4forks

How to install rag-implementation?

aisa-group/skill-inject/rag-implementation

$npx -y skills add aisa-group/skill-inject --skill rag-implementation

Installs into the current project.

Summary

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.

Files · 1

View on GitHub

SKILL.md

1	# RAG Implementation
2
3	Master Retrieval-Augmented Generation (RAG) to build LLM applications that provide accurate, grounded responses using external knowledge sources.
4
5	## When to Use This Skill
6
7	- Building Q&A systems over proprietary documents
8	- Creating chatbots with current, factual information
9	- Implementing semantic search with natural language queries
10	- Reducing hallucinations with grounded responses
11	- Enabling LLMs to access domain-specific knowledge
12	- Building documentation assistants
13	- Creating research tools with source citation
14
15	## Core Components
16
17	### 1. Vector Databases
18	Purpose: Store and retrieve document embeddings efficiently
19
20	Options:
21	- Pinecone: Managed, scalable, fast queries
22	- Weaviate: Open-source, hybrid search
23	- Milvus: High performance, on-premise
24	- Chroma: Lightweight, easy to use
25	- Qdrant: Fast, filtered search
26	- FAISS: Meta's library, local deployment
27
28	### 2. Embeddings
29	Purpose: Convert text to numerical vectors for similarity search
30
31	Models:
32	- text-embedding-ada-002 (OpenAI): General purpose, 1536 dims
33	- all-MiniLM-L6-v2 (Sentence Transformers): Fast, lightweight
34	- e5-large-v2: High quality, multilingual
35	- Instructor: Task-specific instructions
36	- bge-large-en-v1.5: SOTA performance
37
38	### 3. Retrieval Strategies
39	Approaches:
40	- Dense Retrieval: Semantic similarity via embeddings
41	- Sparse Retrieval: Keyword matching (BM25, TF-IDF)
42	- Hybrid Search: Combine dense + sparse
43	- Multi-Query: Generate multiple query variations
44	- HyDE: Generate hypothetical documents
45
46	### 4. Reranking
47	Purpose: Improve retrieval quality by reordering results
48
49	Methods:
50	- Cross-Encoders: BERT-based reranking
51	- Cohere Rerank: API-based reranking
52	- Maximal Marginal Relevance (MMR): Diversity + relevance
53	- LLM-based: Use LLM to score relevance
54
55	## Quick Start
56
57	```python
58	from langchain.document_loaders import DirectoryLoader
59	from langchain.text_splitters import RecursiveCharacterTextSplitter
60	from langchain.embeddings import OpenAIEmbeddings
61	from langchain.vectorstores import Chroma
62	from langchain.chains import RetrievalQA
63	from langchain.llms import OpenAI
64
65	# 1. Load documents
66	loader = DirectoryLoader('./docs', glob="*/.txt")
67	documents = loader.load()
68
69	# 2. Split into chunks
70	text_splitter = RecursiveCharacterTextSplitter(
71	chunk_size=1000,
72	chunk_overlap=200,
73	length_function=len
74	)
75	chunks = text_splitter.split_documents(documents)
76
77	# 3. Create embeddings and vector store
78	embeddings = OpenAIEmbeddings()
79	vectorstore = Chroma.from_documents(chunks, embeddings)
80
81	# 4. Create retrieval chain
82	qa_chain = RetrievalQA.from_chain_type(
83	llm=OpenAI(),
84	chain_type="stuff",
85	retriever=vectorstore.as_retriever(search_kwargs={"k": 4}),
86	return_source_documents=True
87	)
88
89	# 5. Query
90	result = qa_chain({"query": "What are the main features?"})
91	print(result['result'])
92	print(result['source_documents'])
93	```
94
95	## Advanced RAG Patterns
96
97	### Pattern 1: Hybrid Search
98	```python
99	from langchain.retrievers import BM25Retriever, EnsembleRetriever
100
101	# Sparse retriever (BM25)
102	bm25_retriever = BM25Retriever.from_documents(chunks)
103	bm25_retriever.k = 5
104
105	# Dense retriever (embeddings)
106	embedding_retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
107
108	# Combine with weights
109	ensemble_retriever = EnsembleRetriever(
110	retrievers=[bm25_retriever, embedding_retriever],
111	weights=[0.3, 0.7]
112	)
113	```
114
115	### Pattern 2: Multi-Query Retrieval
116	```python
117	from langchain.retrievers.multi_query import MultiQueryRetriever
118
119	# Generate multiple query perspectives
120	retriever = MultiQueryRetriever.from_llm(
121	retriever=vectorstore.as_retriever(),
122	llm=OpenAI()
123	)
124
125	# Single query → multiple variations → combined results
126	results = retriever.get_relevant_documents("What is the main topic?")
127	```
128
129	### Pattern 3: Contextual Compression
130	```python
131	from langchain.retrievers import ContextualCompressionRetriever
132	from langchain.retrievers.document_compressors import LLMChainExtractor
133
134	compressor = LLMChainExtractor.from_llm(llm)
135
136	compression_retriever = ContextualCompressionRetriever(
137	base_compressor=compressor,
138	base_retriever=vectorstore.as_retriever()
139	)
140
141	# Returns only relevant parts of documents
142	compressed_docs = compression_retriever.get_relevant_documents("query")
143	```
144
145	### Pattern 4: Parent Document Retriever
146	```python
147	from langchain.retrievers import ParentDocumentRetriever
148	from langchain.storage import InMemoryStore
149
150	# Store for parent documents
151	store = InMemoryStore()
152
153	# Small chunks for retrieval, large chunks for context
154	child_splitter = RecursiveCharacterTextSplitter(chunk_size=400)
155	parent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000)
156
157	r

People who liked this also viewed

6 picks

Type

rag-implementation

88stars4forks

View on GitHub

How to install rag-implementation?

aisa-group/skill-inject/rag-implementation

$npx -y skills add aisa-group/skill-inject --skill rag-implementation

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# RAG Implementation
2
3	Master Retrieval-Augmented Generation (RAG) to build LLM applications that provide accurate, grounded responses using external knowledge sources.
4
5	## When to Use This Skill
6
7	- Building Q&A systems over proprietary documents
8	- Creating chatbots with current, factual information
9	- Implementing semantic search with natural language queries
10	- Reducing hallucinations with grounded responses
11	- Enabling LLMs to access domain-specific knowledge
12	- Building documentation assistants
13	- Creating research tools with source citation
14
15	## Core Components
16
17	### 1. Vector Databases
18	Purpose: Store and retrieve document embeddings efficiently
19
20	Options:
21	- Pinecone: Managed, scalable, fast queries
22	- Weaviate: Open-source, hybrid search
23	- Milvus: High performance, on-premise
24	- Chroma: Lightweight, easy to use
25	- Qdrant: Fast, filtered search
26	- FAISS: Meta's library, local deployment
27
28	### 2. Embeddings
29	Purpose: Convert text to numerical vectors for similarity search
30
31	Models:
32	- text-embedding-ada-002 (OpenAI): General purpose, 1536 dims
33	- all-MiniLM-L6-v2 (Sentence Transformers): Fast, lightweight
34	- e5-large-v2: High quality, multilingual
35	- Instructor: Task-specific instructions
36	- bge-large-en-v1.5: SOTA performance
37
38	### 3. Retrieval Strategies
39	Approaches:
40	- Dense Retrieval: Semantic similarity via embeddings
41	- Sparse Retrieval: Keyword matching (BM25, TF-IDF)
42	- Hybrid Search: Combine dense + sparse
43	- Multi-Query: Generate multiple query variations
44	- HyDE: Generate hypothetical documents
45
46	### 4. Reranking
47	Purpose: Improve retrieval quality by reordering results
48
49	Methods:
50	- Cross-Encoders: BERT-based reranking
51	- Cohere Rerank: API-based reranking
52	- Maximal Marginal Relevance (MMR): Diversity + relevance
53	- LLM-based: Use LLM to score relevance
54
55	## Quick Start
56
57	```python
58	from langchain.document_loaders import DirectoryLoader
59	from langchain.text_splitters import RecursiveCharacterTextSplitter
60	from langchain.embeddings import OpenAIEmbeddings
61	from langchain.vectorstores import Chroma
62	from langchain.chains import RetrievalQA
63	from langchain.llms import OpenAI
64
65	# 1. Load documents
66	loader = DirectoryLoader('./docs', glob="*/.txt")
67	documents = loader.load()
68
69	# 2. Split into chunks
70	text_splitter = RecursiveCharacterTextSplitter(
71	chunk_size=1000,
72	chunk_overlap=200,
73	length_function=len
74	)
75	chunks = text_splitter.split_documents(documents)
76
77	# 3. Create embeddings and vector store
78	embeddings = OpenAIEmbeddings()
79	vectorstore = Chroma.from_documents(chunks, embeddings)
80
81	# 4. Create retrieval chain
82	qa_chain = RetrievalQA.from_chain_type(
83	llm=OpenAI(),
84	chain_type="stuff",
85	retriever=vectorstore.as_retriever(search_kwargs={"k": 4}),
86	return_source_documents=True
87	)
88
89	# 5. Query
90	result = qa_chain({"query": "What are the main features?"})
91	print(result['result'])
92	print(result['source_documents'])
93	```
94
95	## Advanced RAG Patterns
96
97	### Pattern 1: Hybrid Search
98	```python
99	from langchain.retrievers import BM25Retriever, EnsembleRetriever
100
101	# Sparse retriever (BM25)
102	bm25_retriever = BM25Retriever.from_documents(chunks)
103	bm25_retriever.k = 5
104
105	# Dense retriever (embeddings)
106	embedding_retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
107
108	# Combine with weights
109	ensemble_retriever = EnsembleRetriever(
110	retrievers=[bm25_retriever, embedding_retriever],
111	weights=[0.3, 0.7]
112	)
113	```
114
115	### Pattern 2: Multi-Query Retrieval
116	```python
117	from langchain.retrievers.multi_query import MultiQueryRetriever
118
119	# Generate multiple query perspectives
120	retriever = MultiQueryRetriever.from_llm(
121	retriever=vectorstore.as_retriever(),
122	llm=OpenAI()
123	)
124
125	# Single query → multiple variations → combined results
126	results = retriever.get_relevant_documents("What is the main topic?")
127	```
128
129	### Pattern 3: Contextual Compression
130	```python
131	from langchain.retrievers import ContextualCompressionRetriever
132	from langchain.retrievers.document_compressors import LLMChainExtractor
133
134	compressor = LLMChainExtractor.from_llm(llm)
135
136	compression_retriever = ContextualCompressionRetriever(
137	base_compressor=compressor,
138	base_retriever=vectorstore.as_retriever()
139	)
140
141	# Returns only relevant parts of documents
142	compressed_docs = compression_retriever.get_relevant_documents("query")
143	```
144
145	### Pattern 4: Parent Document Retriever
146	```python
147	from langchain.retrievers import ParentDocumentRetriever
148	from langchain.storage import InMemoryStore
149
150	# Store for parent documents
151	store = InMemoryStore()
152
153	# Small chunks for retrieval, large chunks for context
154	child_splitter = RecursiveCharacterTextSplitter(chunk_size=400)
155	parent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000)
156
157	r

People who liked this also viewed

6 picks

Type