Use LLamaIndexQueryEngine to query Markdown files

Use any LlamaIndex vector store as a Query Engine

This notebook demonstrates the use of the LLamaIndexqueryEngine for retrieval-augmented question answering over documents. It shows how to set up the engine with Docling parsed Markdown files, and execute natural language queries against the indexed data.

The LlamaIndexQueryEngine provides an efficient way to query vectorDBs using any LlamaIndex’s vector store.

We use some Markdown (.md) files as input, feel free to try your own text or Markdown documents.

You can create and add this ChromaDBQueryEngine to DocAgent to use.

%pip install llama-index-vector-stores-chroma==0.4.1
%pip install llama-index==0.12.16
%pip install llama-index llama-index-vector-stores-pinecone==0.4.4

Load LLM configuration

This demonstration requires an OPENAI_API_KEY to be in your environment variables. See our documentation for guidance.

import os

import autogen

config_list = autogen.config_list_from_json(env_or_file="../OAI_CONFIG_LIST")

assert len(config_list) > 0
print("models to use: ", [config_list[i]["model"] for i in range(len(config_list))])

# Put the OpenAI API key into the environment
os.environ["OPENAI_API_KEY"] = config_list[0]["api_key"]

In the first example, we build a LLamaIndexQueryEngine instance using ChromaDB.

Refer to this link for running Chromadb in a Docker container.

from chromadb import HttpClient
from llama_index.vector_stores.chroma import ChromaVectorStore

# we need to set up LlmaIndex's ChromaVectorStore
# Refer to https://docs.llamaindex.ai/en/stable/examples/vector_stores/chroma_metadata_filter/ for more information
chroma_client = HttpClient(
    host="host.docker.internal",
    port=8000,
)

# Option 1: get an existing collection
# use get_collection to get an existing collection
chroma_collection = chroma_client.get_collection("default_collection")

# Option 2: create a new collection
# chroma_collection = chroma_client.create_collection("default_collection")

# Create the Chroma vector store
chroma_vector_store = ChromaVectorStore(chroma_collection=chroma_collection)

Then we can use the LlamaIndex chroma_vector_store to create our AG2 LLamaIndexQueryEngine instance.

from llama_index.llms.openai import OpenAI

from autogen.agentchat.contrib.rag import LlamaIndexQueryEngine

chroma_query_engine = LlamaIndexQueryEngine(
    vector_store=chroma_vector_store,
    llm=OpenAI(model="gpt-4o", temperature=0.0),  # Default model for querying, change if needed
)

Initialize the database with input docs and query it with the engine.

input_dir = (
    "/workspaces/ag2/test/agents/experimental/document_agent/pdf_parsed/"  # Update to match your input directory
)
input_docs = [input_dir + "nvidia_10k_2024.md"]  # Update to match your input documents

# Option 1: initialize the database and add new documents
chroma_query_engine.init_db(new_doc_paths_or_urls=input_docs)

# Option 2: connect to the database without initializing it
# chroma_query_engine.connect_db()

# question = "How much money did Nvidia spend in research and development"
question = "What was the latest quarter's GAAP revenue?"
answer = chroma_query_engine.query(question)
print(answer)

Great, we got the data we needed. Now, let’s add another document.

new_docs = [input_dir + "Toast_financial_report.md"]

chroma_query_engine.add_docs(new_doc_paths_or_urls=new_docs)

And query again from the same database but this time for another corporate entity.

question = "How much money did Toast earn in 2024?"
answer = chroma_query_engine.query(question)
print(answer)

Pinecone

In the second example, we build a similar LLamaIndexQueryEngine instance, but on top of Pinecone.

Refer to https://docs.llamaindex.ai/en/stable/examples/vector_stores/PineconeIndexDemo/ for more details on how to set up Pinecone and PineconeVectorStore

Please put your Pinecone API key in an environment variable called PINECONE_API_KEY.

from pinecone import Pinecone, ServerlessSpec

# Load the Pinecode API key and create the Pinecone object
api_key = os.environ["PINECONE_API_KEY"]
pc = Pinecone(api_key=api_key)

# dimensions are for text-embedding-ada-002, which PineconeVectorStore uses for embedding text by default

# Create an index named ag2
pc.create_index(
    name="ag2",
    dimension=1536,
    metric="euclidean",
    spec=ServerlessSpec(cloud="aws", region="us-east-1"),
)

from llama_index.vector_stores.pinecone import PineconeVectorStore

# Create the vector store
pinecone_index = pc.Index("ag2")
pinecone_vector_store = PineconeVectorStore(pinecone_index=pinecone_index)

pinecone_query_engine = LlamaIndexQueryEngine(
    vector_store=pinecone_vector_store,
    llm=OpenAI(model="gpt-4o", temperature=0.0),  # Default model for querying, change if needed
)

# Initialize the database and add new documents
pinecone_query_engine.init_db(new_doc_paths_or_urls=input_docs)

Query the Pinecone query engine

question = "How much money did Nvidia spend in research and development"
answer = pinecone_query_engine.query(question)
print(answer)

Add another document

new_docs = [input_dir + "Toast_financial_report.md"]
pinecone_query_engine.add_docs(new_doc_paths_or_urls=new_docs)

Query again

question = "How much money did Toast earn in 2024?"
answer = pinecone_query_engine.query(question)
print(answer)

Edit this page