DocChunkerSLM
AI extraction with human-in-the-loop governance.
Trust-grade knowledge from your most complex documents.
Wins throughput - the trust layer is left to you
The governance a regulated enterprise needs before data touches a decision.
Unstructured.io is a real company in the document-ingestion space. This is an independent teardown by Kaushal Khodifad, not affiliated with or endorsed by Unstructured.io, and the comparison reflects publicly documented scope only.
How It Works
Upload
Upload PDF documents
Extract
AI extracts entities
Review
Human validates each extraction
Activate
Approved data enters knowledge store
The governance pipeline
Nothing enters the knowledge store until a human clears it. This is the sandbox, as a flow.
Provenance fails to bind → item is blocked from the approval queue; it can be reviewed but never approved.
Enterprise-Grade Capabilities
AI Extraction
The extraction model extracts entities with a rationale chain and verbatim source quote.
How it's enforcedHide
The extraction model (hosted via OpenRouter in this demo; an in-VPC small model is the v1 target) extracts entities and clauses, returning a rationale chain and a verbatim source quote for every value.
Human-in-the-Loop Review
Claim, inspect provenance, approve or reject - every extraction.
How it's enforcedHide
Every extraction passes through a governed review workflow. Approve/Edit are rejected at the API level (HTTP 409) when provenance is unverified.
Deterministic Provenance
Quoted spans bound to real page + character offsets, server-side.
How it's enforcedHide
The model's quoted span is located inside the parsed PDF text server-side, yielding real page + character offsets. A span that fails to locate is blocked from the approval queue; it can be reviewed but never approved.
Knowledge Activation
Approved entities enter a queryable, auditable knowledge store.
How it's enforcedHide
Approved entities flow into a governed knowledge store - queryable, auditable, and trust-grade certified. Only human-approved values are ever activated.
Multi-Schema Support
Schema templates for six industries; three ship fully bound.
How it's enforcedHide
Schema templates for six industries. The bundled demo corpus ships three fully-bound contracts: a Telecom MSA, a Healthcare BAA, and a Finance credit agreement.
Complete Audit Trail
Every action logged with reviewer, timestamp, and state transition.
How it's enforcedHide
Every action - claim, edit, approve, reject, flag - is recorded with reviewer, timestamp, and state transition. Append-only; no updates or deletes.
Industry Coverage
What is real vs. simulated in this demo
PDF parsing + char offsets - REALdetails
PDF parsing, span location, and every page + character offset are recomputed server-side at request time. The review state machine, provenance gate, no-bulk-approve, and claim ownership are enforced by the API (rejections return HTTP 409/400), not by UI copy. Sandbox extraction calls a live hosted model when a key is configured, and says so when it does not.
Sample contracts fictional - SYNTHETICdetails
The three sample contracts are fictional documents authored for this demo - parties, amounts, and clauses are invented. The curated extractions and confidence values are hand-authored, then verified against the real PDFs at runtime; any that failed verification would be downgraded and blocked from approval. Demo review history is synthesised and ephemeral.
Model class - SYNTHETICdetails
this demo runs a frontier LLM via OpenRouter as a capability stand-in; the v1 architecture targets a fine-tuned small model.
SLM-in-VPC deployment - ROADMAPdetails
SLM-in-VPC deployment (weights and text never leaving the customer perimeter) is the v1 target architecture described in the strategy deck. This portfolio demo runs a hosted frontier model via OpenRouter - the opposite deployment mode, disclosed here deliberately.