The Denodo AI SDK is an open-source component designed to streamline the integration of the Denodo Platform with Large Language Models (LLMs). It provides developers with the essential tools to build high-performance AI agents that can interact natively with enterprise data through a governed data virtualization layer.
By automating the complexities of Retrieval-Augmented Generation (RAG) and VQL query generation, the SDK ensures AI responses are grounded in real-time, factual enterprise context.
- Text-to-VQL: Automatically translates natural language questions into Denodo Virtual Query Language (VQL), allowing AI agents to query the Denodo Platform directly without manual SQL/VQL coding.
- Metadata Search (RAG): Uses vectorization to index and search through your technical and business metadata. This enables agents to find the right data assets instantly using semantic search.
- DeepQuery Agent: A sophisticated research agent that orchestrates complex tasks. It crafts a multi-step execution plan and carries it out by combining metadata discovery with live VQL data extraction.
- Model & Vector Store Agnostic: Provides a flexible architecture to configure and switch between various LLMs and Vector Databases depending on your performance and privacy requirements.
To showcase the AI SDK’s capabilities, a sample chatbot application is included.
The complete user manual for the Denodo AI SDK is available here.
To get started with the AI SDK:
- Clone this repository and
cdinto it - Create a new virtual environment (
python -m venv venv) in the root of the AI SDK's path - Activate the virtual environment (
source venv/bin/activatefor Linux/MacOS or.\venv\Scripts\activatefor Windows) - Install the requirements.txt (
python -m pip install -r requirements.txt) - Rename the configuration templates for both AI SDK (
api/utils/sdk_config.env.example=>api/utils/sdk_config.env) and the sample chatbot (sample_chatbot/chatbot_config.env.example=>sample_chatbot/chatbot_config.env) - Review the configuration files for both the AI SDK and the sample chatbot and configure your own LLM/embeddings providers
The sample chatbot includes a sidebar for specialized chatbots. These chatbots behave like custom agents focused on a specific use case, with their own databases, tags, LLM settings, feature flags, and behavioral constraints.
Specialized chatbots are detected automatically from YAML files placed in api/agents/custom. Each chatbot can also define its own icon by adding an image file to api/agents/custom/icons.
To create a specialized chatbot:
- Create a new YAML file in
api/agents/custom. - Add the chatbot icon file to
api/agents/custom/icons. - Start or restart the sample chatbot application.
The sample chatbot will discover the new configuration and add the chatbot to the sidebar automatically.
For example, you can create api/agents/custom/samples_bank.yaml with:
name: "Banking agent"
id: "banking-agent"
description: "Retail banking analytics assistant."
icon_file_name: "bank_agent.png"
settings:
databases:
- "samples_bank"
llm_provider: "OpenAI"
llm_model: "gpt-4.1-mini"
unstructured_mode: false
feedback_enabled: false
reporting_enabled: false
deepquery_enabled: false
user_edit_llm: falseAnd place the corresponding icon at api/agents/custom/icons/bank_agent.png.
This example creates a specialized chatbot with:
- Access limited to the
samples_bankdatabase - LLM model fixed to
gpt-4.1-mini - Unstructured CSV mode disabled
- Feedback disabled
- Reporting disabled
- DeepQuery disabled
- End users unable to edit the selected LLM
This makes it easy to create focused chatbot experiences for specific teams or workflows while keeping the general chatbot available in parallel.
- A thinking model from either OpenAI/AWS Bedrock/Google Vertex (Not Ollama).
- An minimum allowance of minimum 50RPM OpenAI/AWS Bedrock/Google Vertex.
- Powerful thinking model with over 128k context length.
Depending on your LLM provider, here's a guide on how to configure Denodo DeepQuery:
THINKING_PROVIDER=openai
THINKING_MODEL=gpt-5.2-high
THINKING_PROVIDER = bedrock
THINKING_MODEL = us.anthropic.claude-sonnet-4-20250514-v1:0
AWS_CLAUDE_THINKING = 1
AWS_CLAUDE_THINKING_TOKENS = 2048
Please note that AWS Bedrock requires the previously mentioned extra env variables in sdk_config.env to activate thinking.
THINKING_PROVIDER = google
THINKING_MODEL = gemini-3-pro
GOOGLE_THINKING = 1
GOOGLE_THINKING_TOKENS = 2048
Please note that Google requires the previously mentioned extra env variables in sdk_config.env to activate thinking.
We test our query-to-SQL pipeline on our propietary benchmark across the whole range of LLMs that we support. The benchmark dataset consists of 20+ questions in the finance sector. You may use this benchmark as reference to choose an LLM model.
Latest update: 03/31/2025 on AI SDK version 0.7
| LLM Provider | Model | 🎯 Accuracy | 🕒 LLM execution time (s) | 🔢 Input Tokens | 🔡 Output Tokens | 💰 Cost per Query |
|---|---|---|---|---|---|---|
| OpenAI | GPT-4o | 🟢 | 3.20 | 4,230 | 398 | $0.015 |
| OpenAI | GPT-4o Mini | 🟡 | 4.30 | 4,607 | 445 | $0.001 |
| OpenAI | o1 | 🟢 | 18.60 | 5,110 | 5,438 | $0.403 |
| OpenAI | o1-high | 🟢 | 28.21 | 3,755 | 6,220 | $0.429 |
| OpenAI | o1-low | 🟢 | 15.75 | 3,746 | 2,512 | $0.207 |
| OpenAI | o3-mini | 🟢 | 16.61 | 3,756 | 2,750 | $0.016 |
| OpenAI | o3-mini-high | 🟢 | 28.68 | 3,764 | 8,237 | $0.040 |
| OpenAI | o3-mini-low | 🟢 | 8.66 | 3,811 | 1,080 | $0.009 |
| OpenRouter | Amazon Nova Lite | 🟡 | 1.34 | 4,572 | 431 | <$0.001 |
| OpenRouter | Amazon Nova Micro | 🔴 | 1.29 | 5,788 | 668 | <$0.001 |
| OpenRouter | Amazon Nova Pro | 🟢 | 2.53 | 4,522 | 424 | $0.005 |
| OpenRouter | Claude 3.5 Haiku | 🟢 | 4.38 | 4,946 | 495 | $0.006 |
| OpenRouter | Claude 3.5 Sonnet | 🟢 | 5.02 | 4,569 | 435 | $0.020 |
| OpenRouter | Claude 3.7 Sonnet | 🟢 | 5.46 | 4,695 | 465 | $0.021 |
| OpenRouter | Deepseek R1 671b | 🟢 | 40.28 | 4,138 | 3,041 | $0.011 |
| OpenRouter | Deepseek v3 671b | 🟢 | 13.50 | 4,042 | 424 | $0.005 |
| OpenRouter | Deepseek v3.1 671b | 🟡 | 12.46 | 4,910 | 435 | $0.006 |
| OpenRouter | Llama 3.1 8b | 🔴 | 2.98 | 6,024 | 752 | <$0.001 |
| OpenRouter | Llama 3.1 Nemotron 70b | 🟡 | 9.76 | 5,110 | 892 | $0.001 |
| OpenRouter | Llama 3.3 70b | 🟡 | 10.46 | 4,681 | 402 | $0.001 |
| OpenRouter | Microsoft Phi-4 14b | 🟢 | 6.75 | 4,348 | 728 | <$0.001 |
| OpenRouter | Mistral Small 24b | 🟢 | 5.52 | 5,537 | 563 | <$0.001 |
| OpenRouter | Qwen 2.5 72b | 🟢 | 6.30 | 4,874 | 463 | $0.004 |
| Gemini 1.5 Flash | 🟡 | 2.18 | 4,230 | 398 | <$0.001 | |
| Gemini 1.5 Pro | 🟢 | 5.44 | 4,230 | 398 | $0.007 | |
| Gemini 2.0 Flash | 🟢 | 2.42 | 4,230 | 398 | $0.001 |
Please note that "Input Tokens" and "Output Tokens" is the average input/output tokens per query. Also, each color corresponds to the following range in terms of accuracy:
- 🟢 = 90%+
- 🟡 = 80–90%
- 🔴 = <80%
Finally, any model with its size in the name, i.e.: Llama 3.1 8b, represents an open-source model.
The Denodo AI SDK supports the following LLM providers:
- OpenAI
- AzureOpenAI
- Bedrock
- GoogleAIStudio
- Anthropic
- NVIDIA
- Groq
- Ollama
- Mistral
- SambaNova
- OpenRouter
Where Bedrock refers to AWS Bedrock, NVIDIA refers to NVIDIA NIM and Google refers to Google Vertex AI.
- OpenAI (text-embedding-3-large)
- AzureOpenAI (text-embedding-3-large)
- Bedrock (amazon.titan-embed-text-v2:0)
- Google (text-multilingual-embedding-002)
- Ollama (bge-m3)
- Mistral (mistral-embed)
- NVIDIA (baai/bge-m3)
- GoogleAIStudio (gemini-embedding-exp-03-07)
Where Bedrock refers to AWS Bedrock, NVIDIA refers to NVIDIA NIM and Google refers to Google Vertex AI.
Please see the file called LICENSE.