Intelligent Lab Assistant

An AI-powered assistant for controlling liquid handling robots through natural language commands.

Installation

1. Clone the repository:

git clone https://github.com/lennijusten/intelligent-lab.git
cd intelligent-lab

2. Create and activate a virtual environment (optional but recommended):

python -m venv venv
source venv/bin/activate

3. Install the required packages:

pip install -r requirements.txt

4. Copy the .env.example file to .env and fill in the required environment variables.
5. Set up a LangSmith account for debugging assistance.

Project Structure

• main.py: Entry point for the application.
• src/agent/: Contains the core agent logic.
• state.py: Defines the agent's state structure, including message history, deck state, and code to run.
• prompts.py: Contains prompt templates for various stages of processing.
• models.py: Sets up language models and processing chains.
• tools.py: Defines custom tools like Pydantic classes used by the agent. E.g., defines output schemas for OpenAI function calling.
• graph.py: Configures the agent's workflow as a graph.
• nodes/: Contains individual processing nodes for the agent.
• src/opentrons/: Opentrons-specific implementations.
• src/opentrons/concepts/: Markdown files describing Opentrons-specific concepts.
• configs/: Contains configuration files for the system; currently just the default deck state.

Nodes

• initial_processing: This node processes the user's initial input and generates a structured summary.

• get_info: This node generates a deck state, determines if more information is needed, and either asks the user questions or passes to the next node. The get_info model output conforms to the GetInfoResponse Pydantic class:

  class GetInfoResponse(BaseModel):
      deck_state: DeckState
      info_complete: bool
      questions: str
  

  DeckState is also a Pydantic class with the following structure:

  class DeckState(BaseModel):
      pipettes: Dict[str, str]
      labware: Dict[str, str]
      tip_racks: Dict[str, str]
      modules: Dict[str, str]
  

• concept_finder: This node identifies relevant concepts from a predefined list based on the current message history. For each identified concept, it retrieves associated information from markdown files in /src/opentrons/concepts/. Current concepts include thermocycler_module, temperature_controller_module, transfer, blow_out, discard_tip. The model response follows the RelevantConcepts Pydantic class:

  class RelevantConcepts(BaseModel):
      concepts: List[str]
  

• code_gen: This node generates Python code for the Opentrons API based on the message history.

• code_refinement: This node allows for user feedback and the code output by the code_gen node. Once the user accepts the code, the program finishes.

Dependencies and related projects

• LangChain: Framework for developing applications powered by language models.
• LangGraph: Library for building language agents as graphs.
• Opentrons API: API for controlling Opentrons robots.
• PyLabRobot: (Potential future integration) Hardware-agnostic Python package for controlling liquid handling robots.

Usage

Run the main script:

python main.py --config CONFIG_FILE

To create a summary of the project codebase for an LLM, create a .txt files of filepaths to include and run the create_code_summary.py script:

python create_code_summary.py files_to_include.txt