Co to jest Retrieval-Augmented Generation, czyli RAG?

Przykład kodu w Python który wykorzystuje RAG by przeszukiwać poemat Pan Tadeusz Adama Mickiewicza.

# import
from langchain_community.llms import Ollama
from langchain.chains import RetrievalQA
from langchain_community.embeddings import OllamaEmbeddings
from langchain_community.document_loaders import TextLoader
from langchain_community.vectorstores import Chroma
from langchain_text_splitters import CharacterTextSplitter

# mierzymy czas wykonania, start:
import time
start_time = time.time()

# ładujemy dokument
loader = TextLoader("F:/Ollama/RAG/pan_tadeusz.txt", encoding="utf8")
documents = loader.load()

#  dzielimy go na kawałki (chunks)
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0, separator=".")
docs = text_splitter.split_documents(documents)

# model SBERT - embeddings
#embeddings = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")
# model Ollama - embeddings 
embeddings = OllamaEmbeddings(model="mistral")

# zapisujemy do Chroma ('vector store'), do pliku na dysku nie pamieci
db = Chroma.from_documents(docs, embeddings, persist_directory="F:/RAG/ChromaData")

#zapisujemy rezultat - baze z wektorami
db.persist()

# ile czasu zajelo wykonanie
end_time = time.time()

execution_time = end_time - start_time
print("Execution time:", execution_time, "seconds")

# import
from langchain_community.llms import Ollama
from langchain.chains import RetrievalQA
from langchain_community.embeddings import OllamaEmbeddings
from langchain_community.vectorstores import Chroma

# mierzymy czas wykonania, start:
import time
start_time = time.time()

# reset zmiennej
db=None

embeddings = OllamaEmbeddings(model="mistral")

# odczytaj z bazy Chroma
db = Chroma(persist_directory="F:/Ollama/RAG/data", embedding_function=embeddings)
results = db.get()

llm = Ollama(base_url='http://localhost:11434', model="mistral", temperature=0)

qa_chain = RetrievalQA.from_chain_type(
    llm,
    retriever=db.as_retriever(search_type="similarity", search_kwargs={"k": 3}),
    return_source_documents=True
)

question = "O czym mowa w utworze?"
result = qa_chain.invoke({"query": question})
result["result"]
print(result)

#skasuj wszystko
#db.delete_collection()
#print(results)

# ile czasu zajelo wykonanie
end_time = time.time()

execution_time = end_time - start_time
print("Execution time:", execution_time, "seconds")

Rezultat

Jeśli wykonasz ten skrypt zobaczysz odpowiedz podobną do tej poniżej:

{'query': 'O czym mowa w utworze?', 
'result': " The poem appears to be a narrative about an event or scene involving various noblemen and their possessions. 
There is mention of a judge, a seneschal (Rejent), an assessor, a bernardyn (a friar), Wilbik, Skoluba, Vojski, and Asesor. 
They seem to be discussing the taking away of treasures from churches and the opposition to it. 
There is also mention of Napoleon and his granting of titles and lands to his generals. 
The poem contains descriptions of various objects such as a horse, a ring, golden armor, agun (a f), and a gun (fuzyjka). 
There are mentions of Strapczyna, Wojski, Rejent, and Asesor. They seem to be opposing the taking away of treasures from churches. 
The poem also contains descriptions of objects such as a horse, a judge, Spawnik, and a gun. 
It is called Sagalas London à Bałabanów or Sagalas, which is a famous Polish English English English English poem. 
It speaks of a sword, which was used to crush the animal's body. The poem also mentions a ring, a horse, and a judge. 
They seem to be discussing the taking away of treasures from churches, opposition to it, and their opposition to it."}

Jeśli użyjesz parametru 'return_source_documents=True', w odpowiedzi otrzymasz wypisane 'source_documents' - dokumenty które zostały użyte do wygenerowania odpowiedzi.

Format odpowiedzi

W powyższym skrypcie 'result' ma format 'dictionary' (po polsku 'słownik').  Słowniki służą do przechowywania wartości danych w parach klucz : wartość. Słownik to zbiór uporządkowany*, podlegający zmianom i nie pozwalający na duplikowanie.