Retired terms keep answering
A policy changes, the old PDF stays in the library, and retrieval has no way to know which one is real.
Early access · Interactive preview below
Sourcewell scans your source library for outdated policies, duplicate copies and documents that contradict each other. Owners review each finding and approve the fix, before the assistant repeats the wrong answer.
Exact matches, similarity scores and model suggestions are each labeled for what they are.
Every document has a named owner and a review date, so stale sources surface on schedule.
Sourcewell proposes. A reviewer approves. It never edits a document or an assistant’s knowledge itself.
Each decision is logged with the reviewer, the date and the passages they relied on.
The problem
Support bots, internal copilots and retrieval systems all read from document libraries that people maintain by hand. When those libraries drift, the assistant repeats the drift to customers and employees, confidently.
A policy changes, the old PDF stays in the library, and retrieval has no way to know which one is real.
A finance policy and a support macro disagree. The assistant quotes whichever ranks higher that day.
Copies multiply across folders with no owner, no review date and no record of what was approved.
Today
With Sourcewell
Interactive product preview
This is a working preview of the Sourcewell workspace, running in your browser on sample data. Open a finding, compare the passages, record a decision and watch the library, tasks and history update.
8 open findings · 2 open tasks
Select a bar to open those findings.
open resolved
Fictional company, documents and people. Nothing here is customer data or a measured result, and nothing is saved when you leave the page.
What’s in the first release
Seven capabilities, each one a step in the same review workflow, from the first upload to the signed-off change.
Workflow · Import
Upload document collections, including scanned PDFs, then re-check them nightly, weekly or on your own schedule.
New and changed files are checked without anyone remembering to ask.
Workflow · Import
Every document has a named owner and a full revision trail.
You always know who to ask, and what changed since the last approval.
Workflow · Compare
Identical files are matched by content hash. Passages that say nearly the same thing are scored by embedding similarity.
Copies get retired deliberately, not discovered by a customer.
Workflow · Flag
A language model suggests passages that may conflict and shows both, side by side, with the exact text highlighted.
Reviewers start from the evidence, not a hunch.
Workflow · Flag
Set a review cycle per document and see what is overdue before it becomes a wrong answer.
Old guidance is caught on a schedule, not by accident.
Workflow · Review
Route each finding to the right owner with a due date and track it to resolution.
Findings end in action, not in a shared inbox.
Workflow · Approve
A permanent record of every decision: who approved it, when, and the passages they relied on.
Ready for audits and handovers.
How it works
Architecture: Amazon S3 · Amazon Cognito
Upload document collections. Every file is stored as a versioned record with an owner and a review date.
Sample data
How findings are labeled
Reviewers should know how much weight a flag deserves. Every finding says how it was found.
Identical files are matched by content hash. There is nothing to infer: the bytes match or they don’t.
Passages that say almost the same thing are found with embeddings and shown with their score, side by side.
A language model suggests passages that may conflict and explains why, citing the exact text. It is a lead, not a verdict.
Every flag needs a person to sign off. Sourcewell never changes a document, or an assistant’s knowledge, on its own.
Who it’s for
Technology
Sourcewell’s job is to turn uploaded documents into findings a reviewer can trust, so its architecture follows that path: files land in Amazon S3, Textract and Lambda prepare them, Bedrock supplies similarity and contradiction analysis, and DynamoDB keeps the findings and approval history that the dashboard reads.
Signed-in users upload document collections through authenticated API calls.
Original files and extracted text are kept as versioned objects, one per revision.
Scans are read, files are hashed and split into passages, and scheduled scans start on a timer.
Embeddings find near-duplicate passages. A model proposes potential contradictions with cited text.
Findings, owners, tasks and approval history feed the review dashboard. Pipeline health is monitored.
This describes the architecture Sourcewell is designed around. The interactive preview above runs on sample data in your browser and is not connected to these services.
Business progress and early access
Sourcewell is at the building stage. We’re talking with teams who maintain a knowledge library behind an AI assistant, and we share what exists today, including what doesn’t.
Tell us about your library. We reply from hello@sourcewell.space.
Pricing
Sourcewell is onboarding early-access teams. We confirm plan limits and pricing with each team, so there are no list prices yet.
Pricing on request
Pricing on request
Pricing on request
Plan scopes show the intended structure and may change during early access.
Company
Sourcewell is building infrastructure that helps knowledge managers and AI engineering teams keep document libraries clean, reliable and up to date.
Founder, Sourcewell
Leads Sourcewell’s product direction, focused on the outdated, duplicated and conflicting source documents that undermine AI assistants.
Co-founder, Sourcewell
Leads Sourcewell’s technical architecture and AWS design for document processing, embedding, analysis and finding orchestration.
Contact
Write to us about early access, pricing, security questions or anything else about Sourcewell.