Agents for Question Answering

Overview

This example showcases a hierarchical multi-agent system for question-answering applications. The architecture diagram below shows a supervisor agent that interfaces with the user and dispatches tasks to two worker agents to gather information and come up with answers. The worker RAG agent uses the retrieval tool to retrieve relevant documents from a knowledge base - a vector database. The worker SQL agent retrieves relevant data from a SQL database. Although not included in this example by default, other tools such as a web search tool or a knowledge graph query tool can be used by the supervisor agent to gather information from additional sources. Architecture Overview

The AgentQnA example is implemented using the component-level microservices defined in GenAIComps. The flow chart below shows the information flow between different microservices for this example.

flowchart LR %% Colors %% classDef blue fill:#ADD8E6,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5 classDef orange fill:#FBAA60,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5 classDef orchid fill:#C26DBC,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5 classDef invisible fill:transparent,stroke:transparent; %% Subgraphs %% subgraph DocIndexRetriever-MegaService["DocIndexRetriever MegaService "] direction LR EM([Embedding MicroService]):::blue RET([Retrieval MicroService]):::blue RER([Rerank MicroService]):::blue end subgraph UserInput[" User Input "] direction LR a([User Input Query]):::orchid Ingest([Ingest data]):::orchid end AG_REACT([Agent MicroService - react]):::blue AG_RAG([Agent MicroService - rag]):::blue AG_SQL([Agent MicroService - sql]):::blue LLM_gen{{LLM Service <br>}} DP([Data Preparation MicroService]):::blue TEI_RER{{Reranking service<br>}} TEI_EM{{Embedding service <br>}} VDB{{Vector DB<br><br>}} R_RET{{Retriever service <br>}} %% Questions interaction direction LR a[User Input Query] --> AG_REACT AG_REACT --> AG_RAG AG_REACT --> AG_SQL AG_RAG --> DocIndexRetriever-MegaService EM ==> RET RET ==> RER Ingest[Ingest data] --> DP %% Embedding service flow direction LR AG_RAG <-.-> LLM_gen AG_SQL <-.-> LLM_gen AG_REACT <-.-> LLM_gen EM <-.-> TEI_EM RET <-.-> R_RET RER <-.-> TEI_RER direction TB %% Vector DB interaction R_RET <-.-> VDB DP <-.-> VDB

Why should AI Agents be used for question-answering?

  1. Improve relevancy of retrieved context. RAG agents can rephrase user queries, decompose user queries, and iterate to get the most relevant context for answering a user’s question. Compared to conventional RAG, RAG agents significantly improve the correctness and relevancy of the answer because of the iterations it goes through.

  2. Expand scope of skills. The supervisor agent interacts with multiple worker agents that specialize in different skills (e.g., retrieve documents, write SQL queries, etc.). Thus, it can answer questions with different methods.

  3. Hierarchical multi-agents improve performance. Expert worker agents, such as RAG agents and SQL agents, can provide high-quality output for different aspects of a complex query, and the supervisor agent can aggregate the information to provide a comprehensive answer. If only one agent is used and all tools are provided to this single agent, it can lead to large overhead or not use the best tool to provide accurate answers.

Deploy with docker

1. Set up environment

First, clone the GenAIExamples repo.

export WORKDIR=<your-work-directory>
cd $WORKDIR
git clone https://github.com/opea-project/GenAIExamples.git

Second, set up environment variables.

For proxy environments only
export http_proxy="Your_HTTP_Proxy"
export https_proxy="Your_HTTPs_Proxy"
# Example: no_proxy="localhost, 127.0.0.1, 192.168.1.1"
export no_proxy="Your_No_Proxy"
For using open-source llms
export HUGGINGFACEHUB_API_TOKEN=<your-HF-token>
export HF_CACHE_DIR=<directory-where-llms-are-downloaded> #so that no need to redownload every time
[Optional] OPANAI_API_KEY to use OpenAI models
export OPENAI_API_KEY=<your-openai-key>

Third, set up environment variables for the selected hardware using the corresponding set_env.sh

Gaudi
source $WORKDIR/GenAIExamples/AgentQnA/docker_compose/intel/hpu/gaudi/set_env.sh
Xeon
source $WORKDIR/GenAIExamples/AgentQnA/docker_compose/intel/cpu/xeon/set_env.sh

3. Launch the multi-agent system.

Two options are provided for the llm_engine of the agents: 1. open-source LLMs on Gaudi, 2. OpenAI models via API calls.

Gaudi

On Gaudi, meta-llama/Meta-Llama-3.1-70B-Instruct will be served using vllm. By default, both the RAG agent and SQL agent will be launched to support the React Agent.
The React Agent requires the DocIndexRetriever’s compose.yaml file, so two compose.yaml files need to be run with docker compose to start the multi-agent system.

Note: To enable the web search tool, skip this step and proceed to the “[Optional] Web Search Tool Support” section.

cd $WORKDIR/GenAIExamples/AgentQnA/docker_compose/intel/hpu/gaudi/
docker compose -f $WORKDIR/GenAIExamples/DocIndexRetriever/docker_compose/intel/cpu/xeon/compose.yaml -f compose.yaml up -d
[Optional] Web Search Tool Support
Instructions A web search tool is supported in this example and can be enabled by running docker compose with the `compose.webtool.yaml` file. The Google Search API is used. Follow the [instructions](https://python.langchain.com/docs/integrations/tools/google_search) to create an API key and enable the Custom Search API on a Google account. The environment variables `GOOGLE_CSE_ID` and `GOOGLE_API_KEY` need to be set. 
cd $WORKDIR/GenAIExamples/AgentQnA/docker_compose/intel/hpu/gaudi/
export GOOGLE_CSE_ID="YOUR_ID"
export GOOGLE_API_KEY="YOUR_API_KEY"
docker compose -f $WORKDIR/GenAIExamples/DocIndexRetriever/docker_compose/intel/cpu/xeon/compose.yaml -f compose.yaml -f compose.webtool.yaml up -d

Xeon

On Xeon, only OpenAI models are supported. By default, both the RAG Agent and SQL Agent will be launched to support the React Agent.
The React Agent requires the DocIndexRetriever’s compose.yaml file, so two compose yaml files need to be run with docker compose to start the multi-agent system.

export OPENAI_API_KEY=<your-openai-key>
cd $WORKDIR/GenAIExamples/AgentQnA/docker_compose/intel/cpu/xeon
docker compose -f $WORKDIR/GenAIExamples/DocIndexRetriever/docker_compose/intel/cpu/xeon/compose.yaml -f compose_openai.yaml up -d

4. Ingest Data into the vector database

The run_ingest_data.sh script will use an example jsonl file to ingest example documents into a vector database. Other ways to ingest data and other types of documents supported can be found in the OPEA dataprep microservice located in the opea-project/GenAIComps repo.

cd  $WORKDIR/GenAIExamples/AgentQnA/retrieval_tool/
bash run_ingest_data.sh

Note: This is a one-time operation.

Launch the UI

Open a web browser to http://localhost:5173 to access the UI. Ensure the environment variable AGENT_URL is set to http://$ip_address:9090/v1/chat/completions in ui/svelte/.env or else the UI may not work properly.

The AgentQnA UI can be deployed locally or using Docker. To customize deployment, refer to the AgentQnA UI Guide.

[Optional] Deploy using Helm Charts

Refer to the AgentQnA helm chart for instructions on deploying AgentQnA on Kubernetes.

Validate Services

  1. First look at logs for each of the agent docker containers:

# worker RAG agent
docker logs rag-agent-endpoint

# worker SQL agent
docker logs sql-agent-endpoint

# supervisor agent
docker logs react-agent-endpoint

Look for the message “HTTP server setup successful” to confirm the agent docker container has started successfully.

  1. Use python to validate each agent is working properly:

# RAG worker agent
python $WORKDIR/GenAIExamples/AgentQnA/tests/test.py --prompt "Tell me about Michael Jackson song Thriller" --agent_role "worker" --ext_port 9095

# SQL agent
python $WORKDIR/GenAIExamples/AgentQnA/tests/test.py --prompt "How many employees in company" --agent_role "worker" --ext_port 9096

# supervisor agent: this will test a two-turn conversation
python $WORKDIR/GenAIExamples/AgentQnA/tests/test.py --agent_role "supervisor" --ext_port 9090

How to register other tools with the AI agent

The tools folder contains YAML and Python files for additional tools for the supervisor and worker agents. Refer to the “Provide your own tools” section in the instructions here to add tools and customize the AI agents.