Follow Bank Statement Parser development. Subscribe via RSS or watch the GitHub repository for release notifications.

v0.0.8 — 2026-04-11 (Latest) — "Full Platform"

Closes competitive gaps with a complete platform for bank statement processing.

Added multi-currency balance verification — verify_balance_multi_currency() groups by currency, runs Golden Rule per group.
Added hledger + beancount export — to_hledger() and to_beancount() in bankstatementparser.export.
Added bulk directory scanner — scan_and_ingest() scans folder trees, deduplicates across batch.
Added account mapping rules — AccountMapper with ordered regex rules from JSON config.
Added REST API — FastAPI wrapper with /ingest and /health endpoints ([api] extra).

v0.0.7 — 2026-04-08 — "Universal Vision"

Added direct Ollama bridge (ollama_direct_completion) — bypasses LiteLLM long-prompt hang.
Added strip mode (VisionExtractor.strip_rows=True) — splits dense pages into overlapping bands for small local models.
Changed recommended vision model from llava to minicpm-v (better for OCR/document tasks).

Breaking: Dropped Python 3.9 support (now 3.10–3.14). Python 3.9 reached EOL 2025-10-31.
Added enrichment module (Categorizer, EnrichedTransaction, DEFAULT_CATEGORY_SCHEMA) for LLM-powered transaction categorisation.
Added interactive review mode with --type review CLI command.
Added per-row bounding box extraction (Transaction.source_bbox) for downstream review UIs.
Expanded test suite to 718 tests with 100% branch coverage.

Added hybrid PDF pipeline (smart_ingest()) with deterministic/text-LLM/vision-LLM routing.
Added LLMExtractor for digital PDFs via LiteLLM.
Added VisionExtractor for scanned PDFs via multimodal vision models.
Added Golden Rule balance verification (opening + credits − debits == closing).
Added idempotent deduplication via transaction_hash (MD5 fingerprint).
Added install extras: [hybrid], [hybrid-plus], [hybrid-vision].

Added parallel file parsing with parse_files_parallel() using ProcessPoolExecutor.
Added true streaming for large PAIN.001 files (50 MB+) with bounded memory.
Performance optimisations: CAMT throughput now exceeds 27,000 tx/s, PAIN.001 exceeds 52,000 tx/s.
Added Deduplicator class for detecting exact duplicates and suspected matches with confidence scores.
Added from_string() and from_bytes() methods for in-memory parsing without disk I/O.
Added iter_secure_xml_entries() for secure ZIP archive processing.
Extended CI with performance threshold enforcement.

Added CSV, OFX, QFX, and MT940 parser support.
Added format auto-detection with detect_statement_format() and create_parser().
Added PII redaction (on by default in CLI and streaming mode).
Added export helpers for CSV, JSON, and Excel.
Added optional Polars DataFrame support.
Expanded test suite to 467 tests with 100% branch coverage.

Added PAIN.001 parser (Pain001Parser) for ISO 20022 credit transfer initiation files.
Added CLI interface (python -m bankstatementparser.cli).
Added streaming mode with parse_streaming().
Added input validation and file size limits.

View the full commit history on GitHub.