1~
2Velke jazykove modely (LLM) jako GPT-4 a Claude jsou mimoradne mocne, ale trpi zakladnim omezenim: jejich znalosti jsou zmrazeny v dobe treninku. **Retrieval-Augmented Generation (RAG)** resi presne tento problem kombinaci generativni sily LLM s moznosti ziskavat informace z externich zdroju.
3~
4## Problem: Omezeni LLM
5~
61.  **Staticke znalosti**: LLM vi pouze to, co videl behem treninku.
72.  **Halucinace**: Kdyz LLM nezna odpoved, ma tendenci si ji vymyslet.
83.  **Zadny pristup k soukromym datum**: Genericky LLM nema pristup k interni dokumentaci vasi spolecnosti.
9~
10## Co je RAG?
11~
12```mermaid
13graph LR
14    User["User"] -- "Question" --> Retriever
15    Retriever -- "Search relevant\ndocuments" --> VectorStore["Vector Store"]
16    VectorStore -- "Relevant\ndocuments" --> Retriever
17    Retriever -- "Context + Question" --> LLM
18    LLM -- "Grounded\nresponse" --> User
19```
20~
21## Jak RAG funguje podrobne
22~
23### Faze 1: Indexace
24~
25```mermaid
26graph TD
27    A["Documents\n(PDF, HTML, MD, DB)"] --> B["Document Loader"]
28    B --> C["Text Splitter"]
29    C --> D["Text Chunks"]
30    D --> E["Embedding Model"]
31    E --> F["Numerical Vectors"]
32    F --> G["Vector Store\n(ChromaDB, Pinecone, FAISS)"]
33```
34~
35### Faze 2: Vyhledavani + Generovani
36~
37## Stavba RAG Pipeline s LangChain
38~
39```bash
40pip install langchain langchain-openai langchain-community chromadb
41```
42~
43```python
44from langchain_community.document_loaders import PyPDFLoader, WebBaseLoader, DirectoryLoader, TextLoader
45~
46pdf_loader = PyPDFLoader("docs/manual.pdf")
47pdf_docs = pdf_loader.load()
48all_docs = pdf_docs
49```
50~
51```python
52from langchain.text_splitter import RecursiveCharacterTextSplitter
53~
54text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
55chunks = text_splitter.split_documents(all_docs)
56```
57~
58```python
59from langchain_openai import OpenAIEmbeddings
60from langchain_community.vectorstores import Chroma
61~
62embedding_model = OpenAIEmbeddings(model="text-embedding-3-small")
63vectorstore = Chroma.from_documents(documents=chunks, embedding=embedding_model, persist_directory="./chroma_db")
64```
65~
66```python
67from langchain_openai import ChatOpenAI
68from langchain_core.prompts import ChatPromptTemplate
69from langchain_core.runnables import RunnablePassthrough
70from langchain_core.output_parsers import StrOutputParser
71~
72llm = ChatOpenAI(model="gpt-4o", temperature=0)
73retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 4})
74~
75prompt = ChatPromptTemplate.from_template("""
76Answer the question based only on the provided context.
77If the context does not contain enough information, say you don't know.
78~
79Context:
80{context}
81~
82Question: {question}
83~
84Answer:
85""")
86~
87def format_docs(docs):
88    return "\n\n".join(doc.page_content for doc in docs)
89~
90rag_chain = (
91    {"context": retriever | format_docs, "question": RunnablePassthrough()}
92    | prompt
93    | llm
94    | StrOutputParser()
95)
96```
97~
98## Pokrocile techniky RAG
99~
100Patri sem Multi-Query Retrieval, kontextualni komprese, hybridni vyhledavani a konverzacni RAG s pameti.
101~
102## Osvedcene postupy
103~
1041.  Experimentujte s velikosti fragmentu (500-1500 tokenu).
1052.  Pouzivejte metadata dokumentu.
1063.  Vyhodnocujte kvalitu pomoci frameworku jako [RAGAS](https://docs.ragas.io/).
1074.  Implementujte pipeline pro aktualizaci dokumentu.
1085.  Pridejte re-ranker po pocatecnim vyhledavani.
109~
110## Zaver
111~
112RAG se stal standardni architekturou pro stavbu AI aplikaci s pristupem ke specifickym znalostem. LangChain vyrazne zjednodusuje implementaci.
113~