Sparrow 0.5.0

Structured Data Extraction with Local Vision AI

🚀 What's New

New Model Lineup

The inference model lineup has been redesigned for clarity and performance:

• Standard mode — Ministral 3 14B: fast, reliable extraction for everyday documents
• Tables Only mode — dots.ocr + Sparrow custom processing pipeline: specialized for complex structured table extraction using the dots.ocr → HTML → Sparrow Templates → JSON flow
• Advanced mode — Gemma 4 31B Dense: maximum accuracy for demanding document types

vLLM Backend — NVIDIA RTX 6000 Pro Support

• Production-grade vLLM inference on NVIDIA RTX 6000 Pro (96GB VRAM)
• Mistral Small 3.2 24B running at full precision with 70% GPU utilization
• dots.ocr loaded alongside at 20% GPU utilization via lazy-load + VRAM caching
• VLLM_WORKER_MULTIPROC_METHOD=spawn required to prevent NCCL deadlock when loading two models in the same process

MLX Backend — Apple Silicon Model Updates

Updated Vision LLM support on Mac Mini M4 Pro:

• Ministral 3 14B (replaces Mistral Small 3.2 24B on MLX)
• Gemma 4 31B Dense
• dots.ocr running on MLX for specialized table extraction

Oracle DB 26ai Integration

Sparrow now uses Oracle DB Free 26ai as a relational backend:

• API key management and validation
• Usage tracking and analytics
• Vector Search support
• JSON output storage

Async Task Queue — Celery + Redis

Introduced async task processing:

• Two queues: data_queue, file_queue
• Flower dashboard for monitoring
• Improved throughput for concurrent document processing requests

---

🔧 Improvements

Tables Only Mode — Sparrow Templates Pipeline

The Tables Only mode implements Sparrow's core proprietary pipeline for structured table extraction. dots.ocr processes the document and outputs HTML, which Sparrow Templates then map to a precisely defined JSON schema. This two-stage approach delivers reliable, repeatable extraction from complex financial and operational tables — going well beyond what a Vision LLM alone can achieve with unstructured table layouts.

Field-Level Extraction Hints

Passing hints alongside documents significantly improves LLM accuracy. Hints cover attention steering to footers and fine print, field disambiguation (e.g., VAT number assignment rules), date format normalization, and priority ordering for ambiguous fields. The mechanism works by converting hard implicit tasks into easier explicit ones.

OCR Fallback Pattern

When critical fields return null in default image-only mode due to visual attention blind spots, OCR mode (text + image) is used as a targeted fallback — not a wholesale replacement for default mode.

New Model Accuracy Gains

Ministral 3 14B on MLX and Gemma 4 31B Dense bring measurable accuracy improvements across document types. Mistral Small 3.2 24B on vLLM continues to lead on production deployments. Model rankings differ between MLX and vLLM platforms — a runtime and pipeline difference, not a fundamental model quality difference.

Agent Functionality — Beyond Document Extraction

Sparrow Agents support multi-step workflows that go beyond image-based data extraction. This includes non-image data analysis pipelines — for example, bonds and financial instrument data analysis where structured JSON output from prior extraction steps feeds downstream agent reasoning and decision logic.

---

⚠️ Breaking Changes

• Model configuration keys renamed to reflect new Standard / Tables Only / Advanced naming
• Qwen models removed from all backends
• Oracle DB replaces previous lightweight storage backend — migration required for existing deployments

---

Full Changelog: [0.4.3...0.5.0](https://github.com/katanaml/sparrow/compare/v0.4.3...v0.5.0)
Live instance: [sparrow.katanaml.io](https://sparrow.katanaml.io/)
Commercial licensing & consulting: abaranovskis@redsamuraiconsulting.com