🚀 Quarry: Simplifying LLM Agents with a SQL/Python-First Approach

Quarry is your all-in-one solution for building powerful agents using steampipe with a SQL/Python-First approach. No more wrestling with APIs—just install the right plugin from the Steampipe Hub and start querying.

🔎 Zero ETL → Fetch real-time data from APIs directly into SQL.

🔄 Modular & Extensible → Combine SQL pipelines and Python functions to build intelligent agents.

🧠 Smart Reasoning → Leverage Monte Carlo Tree Search (MCTS) for advanced decision-making.

⚙️ Key Principles

• SQL + Python Only: Simple, clean, and powerful—no extra tools needed.
• Modular Tool Learning: Build and refine Python functions as reusable tools.
• Advanced Reasoning: MCTS with self-evaluation improves code generation and task-solving.
• Human-in-the-Loop: Seamlessly review, correct, and enhance generated tools.

🌱 Optional

Get Better Results with Curriculum Learning

Start simple. Gradually increase task complexity to improve performance and reasoning.

🔄 Share the previous results using CSV plugin

• Use csv plugin.
• Include def Save_results_in_CSV tool with workdir being consistent with path in:

 ~/.steampipe/config/csv.spc

This way you can use the memory of previous results as another table in Postgres/Steampipe

🏗️ Architecture Overview

flowchart TD
    Start([Start]) --> InitLoad[Load Initial Python Functions\ninto Kernel Interpreter]
    InitLoad --> ConnectDB[Connect to Steampipe\nPostgres Database]
    
    subgraph TaskGeneration[Task Generation Process]
        ConnectDB --> HighGoal[Receive High Level Goal]
        HighGoal --> SubTasks[Generate Subtasks\nIncreasing Complexity]
    end
    
    SubTasks --> GetTools[Retrieve Tools\nfrom VectorDB]
    
    subgraph MCTSProcess[MCTS Process]
        GetTools --> MCTSLoop{MCTS\nIterations}
        MCTSLoop --> GenCode[Generate\nModular Code]
        GenCode --> ExecCode[Execute in\nPython Interpreter]
        ExecCode --> ReactPrompt[React Prompting]
        ReactPrompt --> MCTSLoop
        
        %% Add data persistence branch
        ExecCode --> SaveResults{Save Results?}
        SaveResults --> |Yes| SaveCSV[Save as CSV\nin Plugin Folder]
        SaveCSV --> UpdateSteampipe[Auto-Import as\nSteampipe Table]
        UpdateSteampipe --> |New Data Available| ConnectDB
        SaveResults --> |No| ReactPrompt
    end
    
    MCTSProcess --> BFSTrace[BFS over Best Trace]
    
    subgraph ToolLearning[Tool Learning & Environment Sync]
        BFSTrace --> ExtractCode[Extract Code\nfrom Best Trace]
        ExtractCode --> NewTools[Create New\nLearned Tools]
        NewTools --> |Parallel Update| SyncProcess{Synchronous\nUpdate}
        SyncProcess --> UpdateVDB[Add Tools to VectorDB]
        SyncProcess --> UpdateKernel[Add Tools to\nKernel Global Env]
    end
    
    UpdateVDB --> |Tool Ready| GetTools
    UpdateKernel --> |Environment Ready| GetTools

    %% New UI Component for Editing Tools
    subgraph UIComponent[UI/Streamlit Tool Editor]
        EditToolsUI[Streamlit UI\nfor Editing Tools]
        EditToolsUI --> FetchTools[Fetch Tools from\nVectorDB]
        FetchTools --> ModifyTools[Edit/Update/Delete Tools]
        ModifyTools --> UpdateVDB
    end

    GetTools --> EditToolsUI

    style Start fill:#f9f,stroke:#333,stroke-width:4px
    style TaskGeneration fill:#e6f3ff,stroke:#333
    style MCTSProcess fill:#fff0e6,stroke:#333
    style ToolLearning fill:#e6ffe6,stroke:#333
    style UIComponent fill:#fffbe6,stroke:#333

    classDef processNode fill:#f9f,stroke:#333,stroke-width:2px
    classDef decision fill:#FFD700,stroke:#333,stroke-width:2px
    classDef dataNode fill:#90EE90,stroke:#333,stroke-width:2px
    class MCTSLoop,SyncProcess,SaveResults decision
    class GenCode,ExecCode,ReactPrompt,ModifyTools processNode
    class SaveCSV,UpdateSteampipe,FetchTools dataNode

    %% Adding notes for clarity
    note1[Both VectorDB and Kernel\nEnvironment stay synchronized]
    note2[Results become queryable\nby future tasks]
    note3[Streamlit UI enables manual\nediting of tools in VectorDB]
    style note1 fill:#fff,stroke:#333,stroke-width:1px
    style note2 fill:#fff,stroke:#333,stroke-width:1px
    style note3 fill:#fff,stroke:#333,stroke-width:1px

    SyncProcess --> note1
    UpdateSteampipe --> note2
    EditToolsUI --> note3

Loading

🔧 Requirements

1. Install Steampipe

Download and install from steampipe.

2. Install Plugins

Install relevant plugins. For the demo (no API keys required):

steampipe plugin install exec
steampipe plugin install finance
steampipe plugin install csv

Start the service:

steampipe service start --show-password

🚀 Running Quarry

Estimate LLM Cost:

Cost ≈ num_actions * depth_limit * n_iters * 2 LLM calls

• num_actions: Candidates per iteration
• depth_limit: MCTS depth
• 2: For generation + evaluation

env variables (steampipe postgres and LLM keys)

Set your env variable in .env file:

DB_HOST=host.docker.internal if using docker 
DB_PORT=
DB_NAME=
DB_USER=
DB_PASSWORD=
OPENAI_API_KEY=
ANTHROPIC_API_KEY=

Run in Docker for Safety

Modify your goal prompt in src/quarry/local.py (demo code) and run:

sh quarry.sh

Run the Demo without Docker

Install UV Package Manager

Install from uv:

uv venv --python 3.12
uv run demo

🛠️ Inspect & Improve Learned Tools

Not all generated tools will be perfect! Use our Streamlit UI to review, edit, and enhance them:

uv run streamlit run tool-collection-manager.py

🔄 Extend Quarry

Want to expand to other DB? Check out Vanna AI for adapting Quarry to other databases.

🙏 Acknowledgements

• Vanna AI
• LLM Reasoners
• Adobe Dynasaur
• Self-Refine: Improving LLMs with Self-Feedback

🌱 Future releases

• Training codes for finetuning the base LLM models to reduce cost in MCTS search.

⭐️ Show Your Support!

If you find Quarry useful, please star ⭐ this repository and share it! Feedback and contributions are always welcome. 😊