Natural Language to SPARQL translation for the LiITA knowledge base
Project description
NL2SPARQL
Natural Language to SPARQL translation for the LiITA (Linking Italian) linguistic knowledge base.
Overview
NL2SPARQL translates natural language questions (in Italian or English) into SPARQL queries for querying the LiITA knowledge base. It uses a hybrid retrieval system combined with LLM-based query synthesis to generate SPARQL queries.
Features
- Multi-LLM Support: Works with OpenAI, Anthropic, Mistral, Google Gemini, and local Ollama models
- Hybrid Retrieval: Combines semantic search (sentence transformers + FAISS), BM25, and pattern matching
- Ontology-Aware: Semantic search over ontology definitions to discover relevant properties and classes
- Domain-Specific Constraints: Built-in knowledge of LiITA's architecture (emotions, translations, semantic relations)
- Query Validation: Syntax checking, endpoint validation, and constraint verification
- Auto-Fix: Automatically fixes case-sensitive filters and detects variable reuse bugs
- Bilingual: Supports questions in both Italian and English
- Agentic Mode: LangGraph-powered agent with self-correction and ontology exploration
- MCP Server: Model Context Protocol server for integration with Claude Desktop and other MCP clients
- Web UI: Gradio-based web interface for interactive query generation
Installation
# Basic installation
pip install liita-nl2sparql
# With specific LLM provider
pip install liita-nl2sparql[openai] # For OpenAI
pip install liita-nl2sparql[anthropic] # For Anthropic (Claude)
pip install liita-nl2sparql[mistral] # For Mistral AI
pip install liita-nl2sparql[gemini] # For Google Gemini
pip install liita-nl2sparql[ollama] # For local Ollama models
# All providers
pip install liita-nl2sparql[all]
# Agentic mode (LangGraph-based)
pip install liita-nl2sparql[agent-openai] # Agent with OpenAI
pip install liita-nl2sparql[agent-anthropic] # Agent with Anthropic
pip install liita-nl2sparql[agent-all] # Agent with all providers
# MCP server (for Claude Desktop integration)
pip install liita-nl2sparql[mcp-openai] # MCP with OpenAI
pip install liita-nl2sparql[mcp-anthropic] # MCP with Anthropic
pip install liita-nl2sparql[mcp-all] # MCP with all providers
# Web UI (Gradio)
pip install liita-nl2sparql[ui] # Gradio web interface
Development Installation
git clone https://github.com/tonazzog/nl2sparql.git
cd nl2sparql
pip install -e ".[dev,all]"
Configuration
Set your API key as an environment variable:
Linux / macOS:
export OPENAI_API_KEY="your-api-key"
export ANTHROPIC_API_KEY="your-api-key"
export MISTRAL_API_KEY="your-api-key"
export GEMINI_API_KEY="your-api-key"
Windows (Command Prompt):
set OPENAI_API_KEY=your-api-key
set ANTHROPIC_API_KEY=your-api-key
set MISTRAL_API_KEY=your-api-key
set GEMINI_API_KEY=your-api-key
Windows (PowerShell):
$env:OPENAI_API_KEY="your-api-key"
$env:ANTHROPIC_API_KEY="your-api-key"
$env:MISTRAL_API_KEY="your-api-key"
$env:GEMINI_API_KEY="your-api-key"
Ollama runs locally and does not require an API key.
Quick Start
For an interactive tutorial, see the Quick Start Notebook.
Usage
Command Line Interface
# Basic translation (Italian)
nl2sparql translate "Quali lemmi esprimono tristezza?"
# Basic translation (English)
nl2sparql translate "Find all words that express sadness"
# Specify provider and model
nl2sparql translate -p anthropic "What are the hyponyms of vehicle?"
# Save output to file
nl2sparql translate "Definition of love" -o query.sparql
# Verbose output with validation details
nl2sparql translate -V "Find the Sicilian translations of 'house'"
# Validate an existing query
nl2sparql validate query.sparql
# List available models
nl2sparql list-models
# Debug retrieval (see which examples are retrieved)
nl2sparql retrieve "What are the parts of the human body?"
# Agentic mode (self-correcting with LangGraph)
nl2sparql agent "Find all nouns expressing sadness"
nl2sparql agent -p anthropic "Trova aggettivi con traduzioni siciliane"
nl2sparql agent --stream "Complex query with step-by-step output"
nl2sparql agent-viz # Show workflow diagram
Python API
Simple Usage
from nl2sparql import translate
# Italian
result = translate("Quali lemmi esprimono tristezza?")
print(result.sparql)
# English
result = translate("Find all nouns that express joy")
print(result.sparql)
Advanced Usage
from nl2sparql import NL2SPARQL
# Initialize with specific provider
translator = NL2SPARQL(
provider="openai",
model="gpt-4.1",
validate=True,
fix_errors=True,
max_retries=3
)
# Translate a question (Italian or English)
result = translator.translate("Find the Sicilian translations of 'casa'")
# Access results
print(result.sparql) # The generated SPARQL query
print(result.detected_patterns) # Detected query patterns
print(result.confidence) # Confidence score
print(result.validation.is_valid) # Validation status
print(result.validation.result_count) # Number of results from endpoint
# Check if query was auto-fixed
if result.was_fixed:
print(f"Query was fixed after {result.fix_attempts} attempts")
Working with Retrieved Examples
from nl2sparql import NL2SPARQL
translator = NL2SPARQL(provider="openai")
result = translator.translate("Parti del corpo umano")
# See which examples were retrieved for few-shot learning
for ex in result.retrieved_examples:
print(f"Score: {ex.score:.3f}")
print(f"Question: {ex.example.nl}")
print(f"SPARQL: {ex.example.sparql[:100]}...")
Agentic Mode (Recommended for Complex Queries)
The agent uses a LangGraph workflow that can analyze, execute, verify, and self-correct queries:
from nl2sparql.agent import NL2SPARQLAgent
# Initialize with provider and optional API key
agent = NL2SPARQLAgent(
provider="openai", # or "anthropic", "mistral", "gemini", "ollama"
model="gpt-4.1", # optional, uses provider default
api_key="sk-...", # optional, uses environment variable
)
# Translate a question
result = agent.translate(
question="Trova tutti i sostantivi che esprimono tristezza",
language="it",
verbose=True
)
# Access results
print(result["sparql"]) # The generated SPARQL query
print(result["confidence"]) # Confidence score (0-1)
print(result["attempts"]) # Number of generation attempts
print(result["result_count"]) # Results from endpoint execution
print(result["is_valid"]) # Whether validation passed
print(result["detected_patterns"]) # Patterns identified
print(result["refinement_history"]) # Previous failed attempts (if any)
Streaming mode to see each step as it executes:
agent = NL2SPARQLAgent(provider="anthropic")
for node_name, state in agent.stream("Find adjectives with Sicilian translations"):
print(f"[{node_name}] completed")
if node_name == "execute":
print(f" Results: {state.get('result_count', 0)}")
Async support:
import asyncio
async def main():
agent = NL2SPARQLAgent(provider="openai")
result = await agent.atranslate("Trova verbi con emozioni positive")
print(result["sparql"])
asyncio.run(main())
Ontology Retrieval
The agent uses semantic search over an ontology catalog to discover relevant properties and classes. This helps when the LLM needs to find the right vocabulary for a query:
from nl2sparql.retrieval import OntologyRetriever
retriever = OntologyRetriever()
# Find properties related to "broader meaning" (e.g., for hypernyms)
results = retriever.retrieve_properties("broader meaning", top_k=5)
for item in results:
print(f"{item.entry.prefix_local}: {item.entry.short_text}")
# lexinfo:hypernym: A term with a broader meaning
# Format for LLM prompt
prompt_text = retriever.format_for_prompt(results)
The ontology catalog includes classes and properties from OntoLex-Lemon, LexInfo, SKOS, LiLA, ELITA, and other vocabularies used by LiITA.
MCP Server (Claude Desktop Integration)
The MCP (Model Context Protocol) server exposes NL2SPARQL tools to MCP-compatible clients like Claude Desktop. This allows Claude to translate natural language questions to SPARQL queries for the LiITA knowledge base.
Starting the Server
# Start MCP server with default provider (OpenAI)
nl2sparql mcp serve
# Start with specific provider
nl2sparql mcp serve --provider anthropic
nl2sparql mcp serve --provider ollama --model llama3
# Show configuration
nl2sparql mcp config
Claude Desktop Configuration
Add to your claude_desktop_config.json:
{
"mcpServers": {
"nl2sparql": {
"command": "full path to python.exe in virtual environment",
"args": [ "-m", "nl2sparql.mcp", "serve", "--provider", "mistral", "--api-key", "YOUR_KEY" ]
}
}
}
Available MCP Tools
| Tool | Description |
|---|---|
translate |
Full NL-to-SPARQL translation using configured LLM |
infer_patterns |
Detect query patterns from natural language |
retrieve_examples |
Get similar query examples for few-shot learning |
search_ontology |
Search ontology catalog for relevant properties/classes |
get_constraints |
Get domain constraints for detected patterns |
validate_sparql |
Validate SPARQL (syntax, semantic, endpoint) |
execute_sparql |
Execute query against LiITA endpoint |
fix_case_sensitivity |
Auto-fix case-sensitive filters |
check_variable_reuse |
Detect variable reuse bugs |
Python API
import asyncio
from nl2sparql.mcp import NL2SPARQLMCPServer
from nl2sparql.mcp.server import MCPConfig
# Configure and run the server
config = MCPConfig(
provider="anthropic",
model="claude-sonnet-4-20250514",
)
server = NL2SPARQLMCPServer(config)
asyncio.run(server.run())
Web UI (Gradio)
A Gradio-based web interface for interactive SPARQL query generation. This provides a user-friendly way to translate natural language questions without using the command line or writing Python code.
Starting the Web UI
# Start with default provider (Mistral)
python scripts/gradio_app.py
# Start with specific provider
python scripts/gradio_app.py --provider ollama --model llama3
python scripts/gradio_app.py --provider anthropic --api-key YOUR_KEY
# Create a public shareable link
python scripts/gradio_app.py --provider mistral --share
# Custom port
python scripts/gradio_app.py --port 8080
The UI opens at http://localhost:7860 by default.
Features
| Tab | Description |
|---|---|
| Translate | Convert natural language questions to SPARQL with validation and sample results |
| Agent Translate | LangGraph agent with self-correction (analyze → plan → retrieve → generate → execute → verify → refine) |
| Analyze Patterns | Detect query patterns without generating SPARQL |
| Retrieve Examples | Retrieve similar query examples from the dataset |
| Search Ontology | Browse and search classes/properties in the LiITA ontology |
| Execute SPARQL | Run SPARQL queries directly against the LiITA endpoint |
| Fix Query | Auto-fix case-sensitive string filters, and variable reuse issues |
Translate Tab
The standard Translate tab shows:
- Input field for natural language questions
- Generated SPARQL query with syntax highlighting
- Detected patterns and confidence score
- Validation status (syntax, endpoint)
- Sample results from the query
Agent Translate Tab
The Agent Translate tab uses the LangGraph-powered NL2SPARQLAgent which can:
- Analyze the question and detect patterns
- Plan the query structure
- Retrieve similar examples for few-shot learning
- Generate an initial SPARQL query
- Execute and verify the query against the endpoint
- Self-correct if errors occur (up to 3 attempts)
Watch the agent work through each step in real-time with streaming updates.
Web UI (Gradio Agent)
An agent-based Gradio web interface with dual-LLM architecture. Watch in real-time as the orchestrator decides which tools to call and the translator generates SPARQL queries.
Architecture
The app uses two LLMs:
- Orchestrator: Decides which tools to call (can be a smaller/cheaper model)
- Translator: Expert SPARQL generator (used by the
generate_sparqltool)
Starting the Web UI
# Start with default provider (Mistral for both LLMs)
python scripts/gradio_app_agent.py
# Use different providers for orchestrator and translator
python scripts/gradio_app_agent.py \
--orchestrator-provider anthropic --orchestrator-model claude-3-haiku-20240307 \
--translator-provider openai --translator-model gpt-4.1
# With explicit API keys
python scripts/gradio_app_agent.py \
-op mistral -ok "YOUR_MISTRAL_KEY" \
-tp openai -tk "YOUR_OPENAI_KEY"
# Create a public shareable link
python scripts/gradio_app_agent.py --share
# Custom port
python scripts/gradio_app_agent.py --port 8080
The UI opens at http://localhost:7861 by default.
Command-line Options
| Option | Short | Description |
|---|---|---|
--orchestrator-provider |
-op |
Orchestrator LLM provider (default: mistral) |
--orchestrator-model |
-om |
Orchestrator model (uses provider default) |
--orchestrator-api-key |
-ok |
Orchestrator API key (uses env var if not set) |
--translator-provider |
-tp |
Translator LLM provider (default: mistral) |
--translator-model |
-tm |
Translator model (uses provider default) |
--translator-api-key |
-tk |
Translator API key (uses env var if not set) |
--share |
Create a public shareable link | |
--port |
Port to run on (default: 7861) |
Available Tools
The orchestrator can call these tools:
| Tool | Description |
|---|---|
infer_patterns |
Detect query patterns (EMOTION_LEXICON, TRANSLATION, etc.) |
retrieve_examples |
Find similar SPARQL examples for few-shot learning |
search_ontology |
Search for RDF classes and properties |
get_constraints |
Get domain-specific rules for detected patterns |
generate_sparql |
Generate SPARQL using the translator LLM |
validate_sparql |
Validate query (syntax, semantic, endpoint) |
execute_sparql |
Execute query and return results |
fix_query |
Auto-fix common issues (case sensitivity, SERVICE clauses) |
final_answer |
Return the final SPARQL query to the user |
Real-time Tool Calls
The UI streams tool calls as they happen:
- Running indicators show which tool is currently executing
- Completed indicators show results from each tool
- The SPARQL output updates as soon as a query is generated
- Validation results and fixes are shown in real-time
Supported Query Types
| Query Type | Example Question | Description |
|---|---|---|
| Emotion | "Quali lemmi esprimono tristezza?" | Queries ELITA emotion annotations |
| Translation | "Traduzioni siciliane di casa" | Queries dialect translations (Sicilian, Parmigiano) |
| Definition | "Definizione di amore" | Queries CompL-it sense definitions |
| Semantic Relations | "Iperonimi di cane" | Queries hypernyms, hyponyms, meronyms |
| POS Filter | "Trova tutti i verbi" | Filters by part of speech |
| Morphological | "Lemmi che iniziano con 'pre'" | Pattern matching on word forms |
| Compositional | "Tutti gli animali velenosi" | Complex multi-step reasoning |
Project Structure
nl2sparql/
├── scripts/
│ ├── gradio_app.py # Simple Gradio web UI
│ ├── gradio_app_agent.py # Agent-based Gradio UI (dual LLM)
│ └── test_mcp_tools.py # Direct tool testing
├── notebooks/
│ └── quickstart.ipynb # Interactive tutorial
├── __init__.py # Public API
├── __main__.py # Entry point for python -m nl2sparql
├── cli.py # Command-line interface
├── config.py # Configuration management
├── agent/ # Agentic LangGraph workflow
│ ├── __init__.py # Public API (NL2SPARQLAgent)
│ ├── state.py # State definition for workflow
│ ├── nodes.py # Node implementations (analyze, generate, verify, etc.)
│ └── graph.py # LangGraph workflow definition
├── mcp/ # MCP (Model Context Protocol) server
│ ├── __init__.py # Public API (NL2SPARQLMCPServer)
│ ├── __main__.py # Entry point for python -m nl2sparql.mcp
│ ├── server.py # MCP server implementation
│ ├── tools.py # Tool handler implementations
│ └── resources.py # Resource providers
├── constraints/ # Domain-specific prompts and validation
│ ├── __init__.py # Public API for constraints
│ ├── base.py # Core SPARQL patterns and system prompt
│ ├── emotion.py # ELITA emotion constraints
│ ├── translation.py # Dialect translation constraints
│ ├── semantic.py # CompL-it semantic constraints
│ ├── lexical_relation.py # Synonym/antonym constraints
│ ├── multi_entry.py # Multi-entry pattern validation
│ ├── compositional.py # Complex query reasoning
│ └── prompt_builder.py # Dynamic prompt construction
├── retrieval/ # Hybrid retrieval system
│ ├── __init__.py # Public API for retrieval
│ ├── hybrid_retriever.py # Main retriever combining all methods
│ ├── ontology_retriever.py # Semantic search over ontology definitions
│ ├── embeddings.py # Sentence transformers + FAISS
│ ├── bm25.py # BM25 with pattern boosting
│ └── patterns.py # Query pattern inference (keyword + semantic)
├── generation/ # Query synthesis
│ ├── __init__.py # Public API for generation
│ ├── synthesizer.py # Main NL2SPARQL class
│ └── adapters.py # Query adaptation utilities
├── llm/ # LLM provider abstraction
│ ├── __init__.py # Public API for LLM clients
│ ├── base.py # Abstract client interface
│ ├── openai_client.py # OpenAI implementation
│ ├── anthropic_client.py # Anthropic implementation
│ ├── mistral_client.py # Mistral implementation
│ ├── gemini_client.py # Google Gemini implementation
│ └── ollama_client.py # Ollama implementation
├── validation/ # Query validation
│ ├── __init__.py # Public API for validation
│ ├── syntax.py # rdflib syntax validation
│ ├── endpoint.py # SPARQL endpoint validation
│ └── semantic.py # Constraint-based validation
├── evaluation/ # Evaluation framework
│ ├── __init__.py # Public API for evaluation
│ ├── evaluate.py # Test runner and metrics
│ └── batch_evaluate.py # Multi-model comparison
├── synthetic/ # Synthetic data generation
│ ├── __init__.py # Public API for synthetic generation
│ └── generator.py # Training data generator
└── data/
├── sparql_queries_examples.json # Example queries dataset
├── test_dataset.json # Evaluation test cases
└── ontology.json # Ontology catalog (classes & properties)
LiITA Knowledge Base Architecture
The system understands LiITA's multi-source architecture:
- Main LiITA: Lemmas, POS, morphology (
GRAPH <http://liita.it/data>) - ELITA: Emotion annotations (
GRAPH <http://w3id.org/elita>) - Dialect Translations: Sicilian, Parmigiano (via
vartrans:translatableAs) - CompL-it: Senses, definitions, semantic relations (
SERVICE <https://klab.ilc.cnr.it/graphdb-compl-it/>)
Available Models
| Provider | Default Model | Other Models |
|---|---|---|
| OpenAI | gpt-4.1-mini | gpt-5.2, gpt-4.1, gpt-4.1-nano, gpt-4-turbo, gpt-3.5-turbo |
| Anthropic | claude-sonnet-4-20250514 | claude-opus-4-20250514, claude-3-5-haiku-20241022 |
| Mistral | mistral-large-latest | mistral-medium-latest, mistral-small-latest |
| Gemini | gemini-pro | gemini-pro-vision |
| Ollama | llama3 | mistral, codellama, phi3 |
Evaluation
The package includes a test framework for systematic evaluation of single models and batch comparison of multiple models.
Single Model Evaluation
# Full evaluation with default settings
nl2sparql evaluate
# Evaluate with specific provider
nl2sparql evaluate -p anthropic
# Test only single-pattern queries
nl2sparql evaluate -c single_pattern
# Test specific patterns
nl2sparql evaluate --pattern EMOTION_LEXICON --pattern TRANSLATION
# Save results to file (includes generated SPARQL queries)
nl2sparql evaluate -o report.json
Batch Model Comparison
Compare multiple LLM providers and models:
# Quick comparison (GPT-4o-mini vs Claude 3.5 Haiku)
nl2sparql batch-evaluate -p quick
# Compare all OpenAI models
nl2sparql batch-evaluate -p openai -o ./reports
# Compare default models from all providers
nl2sparql batch-evaluate -p all_defaults -c comparison.json
# Custom model selection
nl2sparql batch-evaluate --provider openai --provider anthropic
Available presets:
quick- Fast comparison with smaller modelsopenai- All OpenAI modelsanthropic- All Anthropic modelsmistral- All Mistral modelsall_defaults- Default model from each provider
Python API
from nl2sparql import NL2SPARQL
from nl2sparql.evaluation import (
evaluate_dataset,
print_report,
save_report,
# Batch evaluation
ModelConfig,
run_batch_evaluation,
create_comparison_report,
print_comparison,
PRESETS,
)
# Single model evaluation
translator = NL2SPARQL(provider="openai")
report = evaluate_dataset(translator, language="it")
print_report(report)
save_report(report, "report.json") # Includes generated SPARQL queries
# Batch model comparison
configs = [
ModelConfig("openai", "gpt-4.1", "GPT-4.1"),
ModelConfig("anthropic", "claude-sonnet-4-20250514", "Claude Sonnet"),
]
results = run_batch_evaluation(configs, output_dir="./reports")
comparison = create_comparison_report(results, "comparison.json")
print_comparison(comparison)
Metrics
- Syntax validity: Percentage of queries that parse correctly
- Endpoint success: Percentage of queries that execute without errors
- Component score: Percentage of expected SPARQL components present
- Pattern detection accuracy: How well the system identifies query types
See docs/evaluation.md for detailed documentation.
Synthetic Data Generation
Generate training data for fine-tuning custom LLMs on NL2SPARQL:
# Generate synthetic (NL, SPARQL) pairs
nl2sparql generate-synthetic -o training_data.jsonl
# With options
nl2sparql generate-synthetic -o data.jsonl -n 10 -m 500 -f alpaca
The generator creates validated training pairs by:
- Generating NL variations of seed examples
- Creating pattern combination questions
- Validating all SPARQL against the endpoint
Output formats: jsonl, json, alpaca, sharegpt, hf (HuggingFace)
See docs/synthetic_data.md for detailed documentation.
License
MIT License - see LICENSE for details.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file liita_nl2sparql-0.7.1.tar.gz.
File metadata
- Download URL: liita_nl2sparql-0.7.1.tar.gz
- Upload date:
- Size: 252.8 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.0
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
81b39f09656be46bc594549657da26f21998491240164aa65ccd1b185ef2517c
|
|
| MD5 |
bab25fb50fbcbc5636eb27e812c81597
|
|
| BLAKE2b-256 |
60ebd1defc0b2853554390fbc3e1b4e5e7827ce718115962ea8c3945723b4edb
|
File details
Details for the file liita_nl2sparql-0.7.1-py3-none-any.whl.
File metadata
- Download URL: liita_nl2sparql-0.7.1-py3-none-any.whl
- Upload date:
- Size: 230.7 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.0
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
fc9574f6b7614b5f374bcde546b1cebdf0e1a9859bad2da1b1004d3d8b70bcc4
|
|
| MD5 |
eca3b7964f1fb1121cd4e41aa183d7e1
|
|
| BLAKE2b-256 |
ca6feaafc4a86a37be8af863b842ac768cc003fd5de32318fef0dfbc962ea1c1
|