What Binah can operate

43 schemas across six sectors. Local-first OCR (PyMuPDF · Docling · Tesseract · Vision) with Claude/OpenAI fallback. Every extracted field carries page + bbox + quote + confidence.

faster-whisper transcription, pause-based speaker diarization, LLM key-event extraction (summary · sentiment · topics · key amounts). Cross-modal rule transcript_amount_unverified ties spoken claims to document-extracted values.

Vision OCR + local EXIF + scene detection. Standalone images route through the vision pipeline, not just the scanned-PDF fallback. Workspace renders preview + scene chip + objects + EXIF + GPS map link.

97 deterministic rules across the six sectors. Universal (entity_consistency, math_checks) plus vertical packs (pitch_deck_matches_sec, loan_app_revenue_vs_filed, lease checks, AQL banding for manufacturing). New rule = 1 class.

Every cross-document link rendered as a graph. Per-field bbox + page + quote + confidence. Workspace renders cross-document Bezier curves; mismatches strike through in warm-rust.

DFG discovery across modalities via PM4Py. Documents + voice events on one log through the shared ProcessEvent primitive. Workspace canvas renders the discovered process model.

8-doc loan packet → revenue cross-check vs filed tax · PFS liabilities vs credit bureau · personal bank vs declared income.

10-K + earnings call cross-modal. Flags disclosure asymmetry between filed numbers and what executives said on the call.

CD amortization math · appraisal-vs-sale · Section J/K balance · HUD-1 reconcile.

GL trial-balance check · bank tie-out · debits = credits to the penny before period closes.

BOM ↔ work order ↔ QA report. AQL banding: lot conforms · review for accept/reject · lot REJECTED.

Plain English → READ-ONLY SELECT against three Binah-runtime tables. Three guardrails (system prompt · pre-execution validator · READ ONLY transaction with statement_timeout). SQL returned alongside rows.

PostgresSink auto-creates three tables on first reconcile: binah_extractions, binah_reconcile_findings, binah_evidence_graph. Customer queries directly with psql · DBeaver · Metabase · Superset.

Customer without Postgres queries Binah's Parquet output directly via DuckDB — no Binah infra needed beyond the .parquet files. Lakehouse-shaped, columnar, fast.

Pre-rolled metrics for executives who don't write SQL. Counter cards (projects audited · fired ERROR/WARNING) plus top fired rules plus recent findings table.

12 stable verbs grouped around connect, process, query, and deliver. Internal implementations remain available to Binah without cluttering agent context.

Standard X-API-Key header. /v1/parse · /v1/reconcile · /v1/extract/custom · /v1/webhooks/ingest (HMAC + SSRF-defended) · /v1/sql/ask · /v1/sql/tables. Customer adopts Binah without learning MCP.

4 sinks via the OutputSink protocol: ParquetFileSink · JsonLinesSink · PostgresSink · UnifiedViewSink. Reconcile-completion hook fires registered sinks automatically.

5 sources: S3 watcher · SEC EDGAR client + Celery beat poller · Plaid bank transactions · universal HMAC-signed webhook · direct upload.

5 first-party ingest paths in production. S3-compatible bucket watcher (R2, MinIO, AWS). SEC EDGAR client with autopoll (Celery beat every 15 min). Plaid bank transactions. Universal HMAC-signed webhook that any Zapier / n8n / Make / custom script can post to. Workspace + REST direct upload.

TODAY: any Slack event or Gmail thread reaches Binah via the universal webhook + a Zapier / n8n forwarder. PLANNED: native OAuth adapters that own the polling and event subscription directly (Slack + Gmail are next in queue, ~200-400 LOC each).

TODAY: file changes from any of these can post to the universal webhook via Zapier / n8n. PLANNED: native adapters with OAuth + shared-drive awareness so customers don't run a middleman for high-volume paths.

TODAY: same webhook path — any of these SaaS tools forwards notifications via Zapier / n8n. PLANNED: REST + GraphQL adapters with proper pagination for bulk sync (currently only push-notification volumes work well).

Any upstream that can POST JSON reaches Binah today — Zapier, n8n, Make, custom scripts. HMAC-signed, SSRF-defended, 50MB cap. This is why 'planned' connectors are still USABLE today — the webhook covers any tool with a workflow-automation forwarder.

Simplest path — drag into workspace or POST to /v1/parse from a script. No auth flow needed for the workspace path; API key for /v1.

Binah routes to 20+ models across OCR (Docling, GLM-OCR, Chandra, DeepSeek-OCR, PaddleOCR, Qianfan-VL, Tesseract), vision (Claude · GPT-4o · Gemini planned), transcription (faster-whisper), and LLM extraction + reasoning (Claude · GPT-4o · Qwen · GLM-4 · DeepSeek v3 · Llama 3.3 · Mistral · Gemini planned). Every model declares its license, cost tier, and local-deployability.

Each model is tagged with a cost tier. Local-first routing tries open-weight models on customer hardware before API fallback; cost-sensitive customers stay on the local tier entirely. Premium tiers (Claude Opus, GPT-4o) reserved for high-confidence-demand work.

Different functions need different models. OCR routes by doc-type (handwriting → Chandra; forms → GLM-OCR; PDFs → Docling). LLM extraction routes by accuracy tier and customer allowlist. Reasoning routes to frontier models for cited evidence quality. Right model, right job.

Customers with vendor restrictions (no China-origin models, no US-cloud APIs, etc.) get a per-customer allowlist. A working local-only mode runs Binah end-to-end without any external API call. Required for health, finance, gov, and enterprise deploys.

Customer runs Binah inside their own VPC. Data never leaves their perimeter. Required for health/finance/government deploys. Full local extraction chain ships in the same package.

Local model serving via vLLM-compatible providers (Qwen, DeepSeek, GLM, InternVL, Llama, Mistral, Phi, OLMo). Per-customer allowlist; working local-only mode.

Every modality, every model wrapped behind a Protocol. Swap engines without touching downstream code. Never hard-code a provider.