stindex 1.1.1

Spatiotemporal Index Extraction from Unstructured Text

STIndex - Spatiotemporal Information Extraction

STIndex is a multi-dimensional information extraction system that uses LLMs to extract temporal, spatial, and custom dimensional data from unstructured text. Features end-to-end pipeline with preprocessing, extraction, and visualization.

🌐 Try the Demo Dashboard

Quick Start

Installation

pip install stindex

# Install spaCy language model (required for NER)
python -m spacy download en_core_web_sm

Basic Extraction

# Extract spatiotemporal entities
stindex extract "On March 15, 2022, a cyclone hit Broome, Western Australia."

# Use specific LLM provider
stindex extract "Text here..." --config openai  # or anthropic, hf

End-to-End Pipeline

from stindex import InputDocument, STIndexPipeline

# Create input documents (URL, file, or text)
docs = [
    InputDocument.from_url("https://example.com/article"),
    InputDocument.from_file("/path/to/document.pdf"),
    InputDocument.from_text("Your text here")
]

# Run full pipeline: preprocessing → extraction → warehouse → visualization
pipeline = STIndexPipeline(
    dimension_config="dimensions",
    output_dir="data/output",
)
results = pipeline.run_pipeline(docs)

Schema Discovery (NEW in v0.6.0)

Automatically discover dimensional schemas from Q&A datasets:

from stindex.pipeline.discovery_pipeline import SchemaDiscoveryPipeline

# Discover schema from medical Q&A dataset
discovery = SchemaDiscoveryPipeline(
    questions_path="data/original/mirage/train.jsonl",
    corpus_path="data/original/medcorp/train.jsonl",
    output_path="cfg/discovered_medical_schema.yml",
    n_clusters=10
)
schema = discovery.run()

# Use discovered schema for extraction
pipeline = STIndexPipeline(
    dimension_config="cfg/discovered_medical_schema.yml"
)
results = pipeline.run_pipeline(docs)

Features:

• Domain-agnostic schema discovery from question-answer datasets
• Two-phase approach: cluster-based initial discovery + refinement
• Outputs hierarchy-based dimension configs compatible with extraction pipeline
• Automatic mandatory dimension inclusion (temporal, spatial)

Supported datasets: MIRAGE, MedCorp, HotpotQA, 2WikiMQA, MuSiQue

Python API (Direct Extraction)

from stindex import DimensionalExtractor

# Initialize with default config (cfg/extract.yml)
extractor = DimensionalExtractor()

# Or specify a config
extractor = DimensionalExtractor(config_path="openai")

# Extract entities
result = extractor.extract("March 15, 2022 in Broome, Australia")

# Access results
print(f"Temporal: {len(result.temporal_entities)} entities")
print(f"Spatial: {len(result.spatial_entities)} entities")

# Raw LLM output available for debugging
if result.extraction_config:
    raw_output = result.extraction_config.get("raw_llm_output") if isinstance(result.extraction_config, dict) else result.extraction_config.raw_llm_output
    print(f"Raw output: {raw_output}")

Server Deployment

MS-SWIFT Server (Model Sharding with Tensor Parallelism)

Deploy a single MS-SWIFT server that uses all available GPUs via tensor parallelism:

# Deploy server (auto-detects GPUs by default)
./scripts/deploy_ms_swift.sh

# Stop server
./scripts/stop_ms_swift.sh

# Check logs
tail -f logs/hf_server.log

Configuration (cfg/hf.yml):

• deployment.port: Server port (default: 8001)
• deployment.model: HuggingFace model ID or local path
• deployment.result_path: Directory for inference logs (default: data/output/result)
• deployment.vllm.tensor_parallel_size:
  • auto (default): Auto-detect all available GPUs
  • Or set manually: 1, 2, 4, etc.
• deployment.vllm.gpu_memory_utilization: GPU memory fraction (default: 0.7)

Configuration

STIndex uses a hierarchical configuration structure organized by module:

Preprocessing Configs (cfg/preprocess/)

• chunking.yml: Document chunking strategies

  • strategy: "sliding_window", "paragraph", "element_based", "semantic"
  • max_chunk_size: Maximum tokens per chunk (default: 1500)
  • overlap: Token overlap between chunks (default: 150)

• parsing.yml: Document parsing settings

  • parsing_method: "unstructured" (recommended) or "simple"
  • Format-specific settings for PDF, HTML, DOCX
  • max_file_size_mb: Maximum file size (default: 50MB)

• scraping.yml: Web scraping configuration

  • rate_limit: Seconds between requests (default: 2.0)
  • timeout: Request timeout (default: 30s)
  • cache.enabled: Enable response caching
  • robots.respect_robots_txt: Respect robots.txt rules

Extraction Configs (cfg/extraction/)

Inference Configs (cfg/extraction/inference/)

• extract.yml: Main extraction configuration

  • llm.llm_provider: "hf", "openai", or "anthropic"
  • extraction.enable_cache: Cache extraction results
  • extraction.auto_save: Auto-save to data/output/yyyy-mm-dd/hh-mm-ss.json
  • extraction.min_confidence: Minimum confidence threshold (0.0-1.0)
  • Context-aware extraction settings
  • Post-processing toggles (reflection, OSM context, relative temporal resolution)

• dimensions.yml: Multi-dimensional extraction definitions (hierarchy-based format v0.6.0+)

  • temporal: ISO 8601 normalized dates with 4-level hierarchy (timestamp → date → month → year)
  • spatial: Geocoded locations with 4-level hierarchy (location → city → state → country)
  • event: Optional categorical dimension for event types (disabled by default)
  • entity: Optional categorical dimension for named entities (disabled by default)
  • Each dimension defines: enabled, extraction_type, schema_type, hierarchy, examples
  • Custom dimensions: Add hierarchical dimensions for domain-specific extraction
  • Migration: Use scripts/migrate_dimension_configs.py to convert old field-based configs

• reflection.yml: Two-pass reflection settings

  • enabled: Enable LLM-based quality filtering (default: false)
  • thresholds: Relevance, accuracy, completeness, consistency scores
  • Context-aware reasoning for temporal/spatial consistency checks
  • Quality scoring with configurable weights

• openai.yml: OpenAI API settings

  • model_name: "gpt-4o-mini", "gpt-4o", "gpt-4.1", etc.
  • temperature: Generation temperature (default: 0.0)
  • max_tokens: Maximum output tokens (default: 2048)
  • Requires: OPENAI_API_KEY environment variable

• anthropic.yml: Anthropic Claude API settings

  • model_name: "claude-sonnet-4-5-20250929" (latest)
  • temperature: Generation temperature (default: 0.0)
  • max_tokens: Maximum output tokens (default: 2048)
  • Requires: ANTHROPIC_API_KEY environment variable

• hf.yml: HuggingFace/MS-SWIFT server settings

  • Client config (llm): API endpoint and generation parameters
    • model_name: Model name as reported by server (e.g., "Qwen3-8B")
    • base_url: Server endpoint (e.g., "http://localhost:8001")
    • max_tokens: Maximum tokens per request (default: 32768)
  • Server config (deployment): Model deployment settings
    • model: HuggingFace model ID (e.g., "Qwen/Qwen3-8B")
    • port: Server port (default: 8001)
    • result_path: Inference log directory (null to disable)
    • vllm.tensor_parallel_size: GPU configuration (auto or number)
    • vllm.gpu_memory_utilization: GPU memory fraction (default: 0.7)
    • vllm.max_model_len: Maximum sequence length (default: 32768)

Post-Processing Configs (cfg/extraction/postprocess/)

• spatial.yml: Geocoding and spatial validation

  • geocoder: "nominatim" (free, OSM) or "google" (requires API key)
  • nominatim.rate_limit: Rate limiting (minimum 1.0 seconds for OSM)
  • cache.enabled: Cache geocoding results
  • disambiguation: Context-aware disambiguation settings
  • validation: Geocoding validation (min_confidence, max_distance_km)

• temporal.yml: Temporal normalization

  • format: "iso8601" (default)
  • timezone.default: Default timezone (default: "UTC")
  • relative.handle_relative: Resolve relative dates (e.g., "Monday" → absolute date)
  • ranges.expand_intervals: Expand date ranges to start/end
  • validation: Year range validation (min_year: 1900, max_year: 2100)

Evaluation Config (cfg/extraction/evaluation/)

• evaluate.yml: Evaluation settings
  • dataset.path: Path to evaluation dataset
  • dataset.sample_limit: Limit number of chunks (null = all)
  • llm.llm_provider: LLM provider for evaluation
  • context_aware.enabled: Enable context-aware extraction
  • Post-processing settings for evaluation

Switching LLM Providers

Edit cfg/extraction/inference/extract.yml:

llm:
  llm_provider: hf  # or openai, anthropic

Or specify at runtime:

extractor = DimensionalExtractor(config_path="openai")

Quick Evaluation

# Sequential mode (default)
stindex evaluate

# With specific config
stindex evaluate --llm-config openai

# Limit samples
stindex evaluate --sample-limit 10

Output Structure

Results are organized by dataset and model:

data/output/evaluations/
└── {dataset_name}-{model_name}/
    ├── eval_{timestamp}_{config}.csv         # Detailed results
    └── eval_{timestamp}_{config}.summary.json # Aggregate metrics

TODOs

• Backend server implementation
• Data warehouse integration

License

MIT License