Deelab-Researcher

Deelab aproximation to create Agents capable of retrieve information as best as we can

Concepts:

Corrective augmented generation (CRAG): Improve the robustness of generation keeping in mind that by retrieving from a static and limited corpus of data you could return sub-optimal documents. So you can use web searches as an extension for augmenting the retrieval results.

Why CRAG ?

As mentioned in the paper “corrective Retrieval Augmented Generation” Large Language Models (LLMs) exhibit hallucinations since the accuracy of generated texts cannot be secured by the parametric knowledge they encapsulated.

References

Resources

LangGraph Orchestation and Agent Generation
FastApi to expose an API
AWS-Bedrock for KDB
Redis for the checkpointer
Poetry for dependency management
Anthropic LLM provider

Install

The required python version is python3.12, you will find all the details by checking the pyproject.toml file.
Install poetry by following these instructions
After install poetry you can install the dependencies by running
```
    poetry install
```

Check the params.example.json, because this is an initial version of this tool we only support AWS BEDROCK but you can manually use ANTHROPIC.

    {
        "provider": "AWS",
        "llm_model_id": "<YOUR BEDROCK MODEL ID>",
        "kdb_retriever_params": {
            "kdb_id": "<YOUR KDB ID>",
            "kdb_max_number_of_results": 3,
            "kdb_region": "us-east-1"
        },
        "web_retriever": {
            "enabled": false,
            "urls": [],
            "is_advanced_search": "true",
            "max_number_of_resources": 5
        },
        "checkpointer": {
            "endpoint": "127.0.0.1",
            "port": 6379,
            "db_number": 0,
            "auth_params": {
                "username": "default",
                "password": "<your password>",
                "ssl": true
            }
        }
    }

Anthropic ref "llm_model_id": "claude-3-5-sonnet-20240620"

Provide the values from your AWS Account

Bedrock with Anthropic "model_id": "us.anthropic.claude-3-5-sonnet-20240620-v1:0" (cross regional inference)
Google with Gemini "model_id': "gemini-1.5-pro-002"
Google with Anthropic model_id: "claude-3-5-sonnet@20240620"

Copy the content of params.example.json in params.json then provide the parameters
- You can enable or disable de web search to do this just use the flag enabled=true/false available in the web_retriever

The checkpointer

The configuration of our checkpointer is pretty simple:

The following parameters are mandatory:

endpoint: Redis host
port: Redis port
db_number: integer
auth_params: is optional, provide these parameters for staging, prod environments.

Checkpointer configuration

    "checkpointer": {
        "endpoint": "127.0.0.1",
        "port": 6379,
        "db_number": 0,
        "auth_params": {
            "username": "default",
            "password": "<your password>",
            "ssl": true
        }
    }

Environment

Copy the .env.example into a new file .env and replace the values

    ANTHROPIC_API_KEY="" (If you are using Anthropic)
    REDIS_ENDPOINT="127.0.0.1" (Your redis endpoint running)
    TAVILY_API_KEY="<your-tavily-api-key>"
    LANGSMITH_TRACING="true"
    LANGSMITH_API_KEY="<your langsmith api key>"
    LANGCHAIN_PROJECT="<your-ideal-project-name>"
    # AWS KDB and BEDROCK MODELS
    AWS_PROFILE="<your aws profile>"
    AWS_DEFAULT_REGION="us-east-1"

estebance/agentic-researcher