rag

class besser.bot.nlp.rag.rag.RAG(bot, vector_store, splitter, llm_name, llm_prompt=None, k=4, num_previous_messages=0)

Bases: object

A Retrieval Augmented Generation (RAG) implementation.

A vector stores contains vectorized representations (i.e., embeddings) of chunks of data (text). For a given input query, a retriever gets the k most similar stored embeddings to the input embedding.

This is usually used to, given a query or question, retrieve those chunks that could be helpful to give it an answer. Then, an LLM is in charge of generating that answer, given the original query and the retrieved data as context.

Parameters:

• bot (Bot) – the bot the RAG engine belongs to

• vector_store (langchain_core.vectorstores.base.VectorStore) – the vector store of the RAG engine

• splitter (langchain_text_splitters.base.TextSplitter) – the text splitter of the RAG engine

• llm_name (str) – the name of the LLM of the RAG engine. It must have been previously created and assigned to the bot

• llm_prompt (str) – the prompt containing the detailed instructions for the answer generation by the LLM. If none is provided, the default prompt will be used

• k (int) – number of chunks to retrieve from the vector store

• num_previous_messages (int) – number of previous messages of the conversation to add to the LLM prompt context. Necessary a connection to MonitoringDB.

_nlp_engine

the NLPEngine that handles the NLP processes of the bot the RAG engine belongs to

Type:

NLPEngine

vector_store

the vector store of the RAG engine

Type:

langchain_core.vectorstores.base.VectorStore

splitter

the text splitter of the RAG engine

Type:

langchain_text_splitters.base.TextSplitter

llm_name

the name of the LLM of the RAG engine. It must have been previously created and assigned to the bot

Type:

str

llm_prompt

the prompt containing the detailed instructions for the answer generation by the LLM. If none is provided, the default prompt will be used

Type:

str

k

number of chunks to retrieve from the vector store

Type:

int

num_previous_messages

number of previous messages of the conversation to add to the LLM prompt context. Necessary a connection to MonitoringDB.

Type:

int

DEFAULT_LLM_PROMPT = "You are an assistant for question-answering tasks. Based on the previous messages in the conversation (if provided), and additional context retrieved from a database (if provided), answer the user question. If you don't know the answer, just say that you don't know. Note that if the question refers to a previous message, you may have to ignore the context since it is retrieved from the database based only on the question (the retrieval does not take into account the previous messages). Use three sentences maximum and keep the answer concise"

create_prompt(history, docs, question, llm_prompt=None)

Creates the prompt for the LLM answer generation.

Parameters:

• history (list[Message]) – the chat history

• docs (list[langchain_core.documents.base.Document]) – the retrieved documents to use as context in the prompt

• question (str) – the user question

• llm_prompt (str) – the prompt containing the detailed instructions for the answer generation by the LLM. If none is provided, the RAG’s default value will be used

Returns:

the LLM prompt

Return type:

str

load_pdfs(path)

Load PDF files from a given location into the RAG’s vector store.

Parameters:

path (str) – the path where the files are located

Returns:

the number of chunks added to the vector store

Return type:

int

run(session=None, message=None, llm_prompt=None, llm_name=None, k=None, num_previous_messages=None)

Run the RAG engine.

Parameters:

• session (Session) – the session of the user that started this request. Must be provided if the chat history wants to be added as context to the LLM prompt. Mandatory if message argument is None (in this case, the last user message will be used as RAG query instead)

• message (str) – the message to be used as RAG query. Mandatory if session argument is None

• llm_prompt (str) – the prompt containing the detailed instructions for the answer generation by the LLM. If none is provided, the RAG’s default value will be used

• llm_name (str) – the name of the LLM to use. If none is provided, the RAG’s default value will be used

• k (int) – the number of (top) documents to get. If none is provided, the RAG’s default value will be used

• num_previous_messages (int) – number of previous messages of the conversation to add to the LLM prompt context. If none is provided, the RAG’s default value will be used. Necessary a connection to MonitoringDB.

Returns:

the resulting RAG message

Return type:

RAGMessage

run_retrieval(question, k=None)

Run retrieval. Given a query, return the k most relevant documents (i.e., chunks) from the RAG’s vector store.

Parameters:

• question (str) – the input query

• k (int) – the number of (top) documents to get. If none is provided, the RAG’s default value will be used

Returns:

the retrieved documents

Return type:

list[langchain_core.documents.base.Document]

class besser.bot.nlp.rag.rag.RAGMessage(llm_name, question, answer, docs)

Bases: object

The result of a RAG execution.

Parameters:

• llm_name (str) – the name of the LLM that generated the answer

• question (str) – the original question

• answer (str) – the generated answer

• docs (list[langchain_core.documents.base.Document]) – the list of documents that were used as additional context for the answer generation

llm_name

the name of the LLM that generated the answer

Type:

str

question

the original question

Type:

str

answer

the generated answer

Type:

str

docs

the list of documents that were used as additional context for the answer generation

Type:

list[langchain_core.documents.base.Document]

to_dict()

Returns a dictionary containing the attributes of the RAGMessage object.