Company Researcher Agent

Company Researcher Agent searches the web for information about a user-supplied company and returns it in a structured format defined by user-supplied JSON schema.

🚀 Quickstart with LangGraph server

Set API keys for the LLM of choice (Anthropic is set by default in src/agent/graph.py) and Tavily API:

cp .env.example .env

Clone the repository and launch the assistant using the LangGraph server:

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/langchain-ai/company-researcher.git
cd company-researcher
uvx --refresh --from "langgraph-cli[inmem]" --with-editable . --python 3.11 langgraph dev

How it works

Company Researcher Agent follows a multi-step research and extraction workflow that separates web research from schema extraction, allowing for better resource management and comprehensive data collection:

• Research Phase: The system performs intelligent web research on the input company:
  • Uses an LLM to generate targeted search queries based on the schema requirements (up to max_search_queries)
  • Executes concurrent web searches via Tavily API, retrieving up to max_search_results results per query
  • Takes structured research notes focused on schema-relevant information
• Extraction Phase: After research is complete, the system:
  • Consolidates all research notes
  • Uses an LLM to extract and format the information according to the user-defined schema
  • Returns the structured data in the exact format requested
• Reflection Phase: The system evaluates the quality of extracted information:
  • Analyzes completeness of required fields
  • Identifies any missing or incomplete information
  • Generates targeted follow-up search queries if needed
  • Continues research until information is satisfactory or max reflection steps reached

Configuration

The configuration for Company Researcher Agent is defined in the src/agent/configuration.py file:

• max_search_queries: int = 3 # Max search queries per company
• max_search_results: int = 3 # Max search results per query
• max_reflection_steps: int = 1 # Max reflection steps

Inputs

The user inputs are:

* company: str - A company to research
* extraction_schema: Optional[dict] - A JSON schema for the output
* user_notes: Optional[str] - Any additional notes about the company from the user

If a schema is not provided, the system will use a default schema (DEFAULT_EXTRACTION_SCHEMA) defined in src/agent/state.py.

Schemas

⚠️ WARNING: JSON schemas require title and description fields for extraction. ⚠️ WARNING: Avoid JSON objects with nesting; LLMs have challenges performing structured extraction from nested objects. See examples below that we have tested.

Here is an example schema that can be supplied to research a company:

• See the trace here.

  Example schema

  {
      "title": "CompanyInfo",
      "description": "Basic information about a company",
      "type": "object",
      "properties": {
          "company_name": {
              "type": "string",
              "description": "Official name of the company"
          },
          "founding_year": {
              "type": "integer",
              "description": "Year the company was founded"
          },
          "founder_names": {
              "type": "array",
              "items": {"type": "string"},
              "description": "Names of the founding team members"
          },
          "product_description": {
              "type": "string",
              "description": "Brief description of the company's main product or service"
          },
          "funding_summary": {
              "type": "string",
              "description": "Summary of the company's funding history"
          }
      },
      "required": ["company_name"]
  }
  

Here is an example of a more complex schema:

• See the reflections steps in the trace here.

  Example complex schema

  HARD_EXTRACTION_SCHEMA = {
      "title": "CompanyInfo",
      "description": "Comprehensive information about a company with confidence tracking",
      "type": "object",
      "properties": {
          "company_name": {
              "type": "string",
              "description": "Official name of the company"
          },
          "verified_company": {
              "type": "boolean",
              "description": "Confirmation this is the intended company, not a similarly named one"
          },
          "similar_companies": {
              "type": "array",
              "items": {"type": "string"},
              "description": "List of similarly named companies that could be confused with the target"
          },
          "distinguishing_features": {
              "type": "string",
              "description": "Key features that distinguish this company from similarly named ones"
          },
          "key_executives": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "name": {"type": "string"},
                      "title": {"type": "string"},
                      "verification_date": {"type": "string"},
                      "confidence_level": {
                          "type": "string",
                          "enum": ["high", "medium", "low", "uncertain"]
                      },
                      "source": {"type": "string"}
                  }
              }
          },
          "org_chart_summary": {
              "type": "string",
              "description": "Brief description of organizational structure"
          },
          "leadership_caveats": {
              "type": "string",
              "description": "Any uncertainties or caveats about leadership information"
          },
          "main_products": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "name": {"type": "string"},
                      "description": {"type": "string"},
                      "launch_date": {"type": "string"},
                      "current_status": {"type": "string"}
                  }
              }
          },
          "services": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "name": {"type": "string"},
                      "description": {"type": "string"},
                      "target_market": {"type": "string"}
                  }
              }
          },
          "recent_developments": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "date": {"type": "string"},
                      "title": {"type": "string"},
                      "summary": {"type": "string"},
                      "source_url": {"type": "string"},
                      "significance": {"type": "string"}
                  }
              },
              "description": "Major news and developments from the last 6 months"
          },
          "historical_challenges": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "issue_type": {"type": "string"},
                      "description": {"type": "string"},
                      "date_period": {"type": "string"},
                      "resolution": {"type": "string"},
                      "current_status": {"type": "string"}
                  }
              },
              "description": "Past challenges, issues, or controversies"
          },
          "sources": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "url": {"type": "string"},
                      "title": {"type": "string"},
                      "date_accessed": {"type": "string"},
                      "information_type": {
                          "type": "array",
                          "items": {"type": "string"},
                          "description": "Types of information sourced from this link (e.g., leadership, products, news)"
                      }
                  }
              }
          },
          "company_summary": {
              "type": "string",
              "description": "Concise, dense summary of the most important company information (max 250 words)"
          }
      },
      "required": [
          "company_name",
          "verified_company",
          "company_summary",
          "key_executives",
          "main_products",
          "sources"
      ]
  }
  

Evaluation

Prior to engaging in any optimization, it is important to establish a baseline performance. This repository includes:

1. A dataset consisting of a list of companies and the expected structured information to be extracted for each company.
2. An evaluation script that can be used to evaluate the agent on this dataset.

Set up

Make sure you have the LangSmith CLI installed:

pip install langsmith

And set your API key:

export LANGSMITH_API_KEY=<your_langsmith_api_key>
export ANTHROPIC_API_KEY=<your_anthropic_api_key>

Evaluation metric

A score between 0 and 1 is assigned to each extraction result by an LLM model that acts as a judge.

The model assigns the score based on how closely the extracted information matches the expected information.

Get the dataset

Create a new dataset in LangSmith using the code in the eval folder:

python eval/create_dataset.py

Run the evaluation

To run the evaluation, you can use the run_eval.py script in the eval folder. This will create a new experiment in LangSmith for the dataset you created in the previous step.

python eval/run_eval.py --experiment-prefix "My custom prefix" --agent-url http://localhost:2024