sangrep

Private beta

AI answers you can audit

sangrep reads the whole bundle, attachments and all, then answers with citations that click straight to the source. Checked before you ever see them.

sangrepmeridian-submission
ContextQ3 vendor due-diligence · Meridian acquisition

Vendor agreement

3.2 · vendor-agreement.pdf · p4

This agreement is entered into by Meridian Systems and the counterparty identified in Schedule A, incorporating each appendix by reference.

§4.2 Either party may terminate with ninety (90) days' written notice.

Fees are set out in the pricing schedule (3.2.1) and reviewed annually each January under section 5.

Chatscope: §3 + annexes

What's the termination notice period?

90 days' written notice, per §4.2 of the vendor agreement.

When do the fees change?

Reviewed annually each January, per the pricing schedule.

Draft a risk summary of the appendix for the deal file.

Draft report

vendor-risk-summary.md

4 findings · every claim cited

Ask the appendix…⌘⏎
214 files parsed · citations validated

Citations that click back

Every citation is checked against your actual documents before it reaches you.

Termination requires 90 days' written notice §4.2 · 3.2 and fees are reviewed annually 3.2.1

§9.9rejected: no such source in scope

Scoped by design

The model sees exactly the part you choose. Nothing else is visible to it.

1 · Overview

2 · Financials

3 · Appendices

3.2 · vendor-agreement.pdf

4 · Correspondenceinvisible

Unpacks everything

Archives inside memos inside emails. The whole package becomes one tree.

meridian-submission.docx
3.2 · vendor-agreement.pdf
3.2.1 · pricing.xlsx

Citations that last

Every part of a document keeps its identity across re-parses. Old citations still land where they should.

node·3.2#a41f

↓ re-parse

node·3.2#a41f

Everything on the record

Every parse, decision, and model call is logged. Records get added, never rewritten.

+ 11:02 summarize · model

+ 11:10 override · you

supersedes ↑ · kept forever

The user flow

  1. Set the context

    Q3 vendor due-diligence
  2. Drop the bundle

    meridian-submission.zip · 214 files
  3. Watch it open up

    1,082 addressable nodes
  4. Ask the part that matters

    ask §3 · Appendices
  5. Click straight to the source

    3.2 · p4 → source

Formats

Whole matters, not single files

A project holds as many files as the case needs, in any mix of formats, nested to any depth: archives inside memos inside archives. Anything sangrep can't parse yet still lands in the tree, flagged, for you to handle.

Parsed todayDOCX packagesPDF, embedded files includedZIP archivesPNG · JPEGPlain textMarkdown
On the roadmapXLSXPPTXEML email threads

New formats ship only when they parse all the way down, embedded files and all.

Current capabilities

What sangrep actually does

Every file, one tree

Upload a bundle and see everything inside it, including the spreadsheet inside the PDF inside the memo. Nothing gets lost on the way in.

You decide how files are treated

For every embedded file you choose: keep it whole, summarize it, replace it with your own note, or leave it out. Your call, on the record.

Documents read like documents

Files open as formatted pages, not walls of extracted text. What you read is exactly what the model reads.

Point, then ask

Scope a chat to the whole bundle, one appendix, or a single clause. Answers draw only from what you picked.

Answers with receipts

Every claim links to its source and gets checked against the tree first. Click through and read the original yourself.

Nothing off the record

Every step is logged as it happens. Ask why the answer says what it says, six months later, and get the whole story.

Knows what it's after

Tell sangrep what the package is about before it reads. It carries that context into every parse and answer, so it knows a schedule from a signature.

Sees inside images

Charts, scans, and signatures are read by an image skill and cited by region, not skipped as pixels.

Reads the links too

It follows references out to the open web, and hands you anything behind a login before it keeps going.

Built to clear enterprise reviewSOC 2 Type II · underwayISO 27001 · underwayGDPRCCPAYour documents never train anyone's models.

Human and AI, together

You decide how much to touch

Auditing comes down to judgment calls, and some of those stay yours. Trim a file out of scope, summarize a section before it goes to the model, override what the parser found. Every call is on the record.

Or flip on autopilot and let sangrep make the routine calls for you, then check its work. You choose the mix, section by section. It is your review either way.

You decideAutopilot

annexes.zip · 12 files

exhibit-a.pdfmaterial to the dealkeep
boilerplate-terms.pdf40 pages, low signalsummarize
blank-cover.pdfnothing to readdrop
exhibit-b.pdfwaiting for your call

3 calls made by you · every one on the record

Autopilot · reviewing packageanalyzing
exhibit-a.pdfmaterial to the dealkeep
boilerplate-terms.pdf40 pages, low signalsummarize
blank-cover.pdfnothing to readdrop
AI decisionsummarize

boilerplate-terms.pdf

40 pages, low signalscope checked · recorded

3 routine calls made · ready for review

The engine

It reasons over a graph, it does not just search

A question does not fetch the nearest paragraph. It travels the structure to the exact node that answers it, and brings the receipt back.

AskTermination notice in the vendor agreement?
meridian-submission.docx
3 · Appendices
3.1 · schedule-a.pdf
3.2 · vendor-agreement.pdfmatch
3.2.1 · pricing.xlsx
annexes.zip
90 days' written notice. §4.2 · 3.2 · p4

Agentic retrieval, not vanilla RAG

Not a single vector lookup and a prayer. sangrep walks the document graph, follows references across files, and pulls in exactly what a question needs. The latest reasoning models make that navigation reliable.

Room for the whole package

The same trick coding agents use on huge repositories. sangrep keeps the whole package in reach and pages in the right sections on demand, so a 5,000-page bundle answers like a short memo.

The right tool per job

One skill reads tables, another parses an image, another follows a link. sangrep reaches for the right tool per node instead of forcing everything through one flattener.

Skills

Not everything is text

An image is evidence and a link is a lead. Dedicated skills read them and cite them like any clause in the document.

Reads what's in the picture

It reads the figure, points to the exact region, and cites the box, so a chart or a signature counts as evidence like anything else.

figure-2 · exposure-by-quarter.pngregion · fig-2

Follows a link, hands off a login

Open pages it reads itself. Anything behind a login it hands back to you, then picks up where it left off with the extra context.

sec.gov/…/10-Kread
portal.meridian.co/deal-roomneeds your login
you logged in · reading resumed+2 pages added

Under the hood

What happens between upload and answer

Reading files is the easy part. This is the machinery that makes the answers safe to use.

01 · intake

The whole mess, as is

meridian-submission.docx3.2 vendor-agreement.pdfpricing.xlsx, 3 deepannexes.zip · 12 insidesoon · SQL, data lakes, Google Docs

nothing gets lost on the way in

02 · identity

Every piece gets an address

§3 Appendices#f21c3.2 agreement#a41f3.2.1 pricing#08d3

citations can outlive re-parses

03 · review

You call the messy files

annexes.zipkeepsummarizedrop

your judgment, on the record

04 · ask

It sees your scope, nothing else

§1 Overview§3 Appendices → the model§4 Correspondence

the model can't wander

05 · verify

Every citation hits a checkpoint

3.2 · p4✓ real§9.9✗ bounced3.2.1✓ real

made-up sources never reach you

The record+ 10:41 parse · 214 files+ 10:44 keep · annexes.zip+ 10:52 ask · §3+ 10:52 verify · 2 ✓ · 1 bouncedEvery step lands here. Nothing is ever rewritten.

You never touch any of this. You drop files, pick a scope, and ask. The pipeline is why the answer comes back with receipts instead of vibes.

Roadmap

Chat is just the front door

Chat came first. The same engine that checks every citation is about to do a lot more.

Agents that work the package

Coming soon

Like the agent panel in your IDE, but pointed at a document bundle. A summarizer takes the appendix while an extractor works the contract, in parallel, each scoped and logged.

Summarizer · §3 Appendicesdone ✓

Term extractor · 3.2 vendor-agreement.pdfrunning

Figure checker · §2 Financialsqueued

Reports with receipts

Coming soon

Turn a package into a draft report where every claim carries its citation. Review it, click the receipts, send it on.

Draft report

Vendor risk summary

Termination requires 90 days' notice §4.2 · 3.2 and fees are reviewed each January 3.2.1

Every claim cited

Plugs into your stack

Coming soon

An MCP server that brings the tree, scoped asks, and citation checks to the tools you already use: IDE assistants, chat platforms, internal agents.

IDE assistantchatinternal agents
MCPsangrep · citations validated

A graph across every source

Coming soon

The tree becomes a knowledge graph that follows references across contracts, appendices, spreadsheets, and links. Ask once, then see the path it took, source by source.

What changes our renewal risk?
§4.2 · 3.2pricing.xlsx · B14security policy · link
90-day notice + annual upliftanswer
§4.2B143 sources linked

A library for repeat work

Coming soon

Save the agent, prompt, or plugin that solved a review once, then run it again on the next package. Start with a team library, then open into a marketplace for workflows teams can trust.

Team marketplacereview workflows

Vendor onboarding review

1 agent3 prompts2 plugins
Add Added
Added to your libraryready to run

Review it together

Coming soon

Shared workspaces for teams: one tree, live cursors, and everyone's calls on the same record. No more findings lost in email threads.

§4.2 Either party may terminate with ninety (90) days' written notice, provided that...

Fees are set out in the pricing schedule and reviewed annually.

Your cloud, your models

Coming soon

An enterprise deployment that lives inside your walls: sangrep runs in your infrastructure, your pipelines feed it, and the models you've already approved do the answering.

Your VPC

your pipelines
sangrepyour LLM
cited answers · on the record

Nothing leaves your network

Beyond files on a disk

Coming soon

Point sangrep at a whole system of record, not just an upload: SQL databases, data lakes, and browser-native docs like Google Docs that never had a filetype.

SQLdata lakeGoogle DocsSharePoint
one addressable tree

Always in sync

Coming soon

Connected sources stay live. sangrep re-reads on change and keeps every version, so you can see exactly what moved between two drafts and when.

vendor-agreement.pdfv3

+2 clauses-1 clausesince v2

FAQ

Questions you probably have

Is this just another chat-with-your-docs tool?
Chat is the first interface, not the whole product. sangrep is a document runtime that keeps structure, enforces scope, and validates citations. Agents, generated reports, and MCP integrations are being built on that same foundation.
How does sangrep compare to chatbots and parsing APIs?
Short version: those are building blocks, sangrep is the finished workspace. Long version:
DimensionChat assistantsChatGPT, Copilot, chat-with-PDF toolsParsing APIsReducto, Unstructured, LlamaParsesangrepthe document runtime
Your documents becomeOne flattened blob of textMarkdown and JSON for your pipelineA stable, addressable tree
Embedded filesIgnored or silently droppedExtracted; handling is up to youInventoried; you decide how each is treated
CitationsThe model's word for itGrounding data, if you wire it upChecked against the tree, rejected if wrong
Scope controlThe prompt is the scopeBuild it yourselfEnforced by the system
Audit trailNoneYour jobAppend-only, built in
Built forQuick answersEngineering teams building pipelinesReviewers who need answers that hold up
What happens to the files I upload?
They stay in your workspace. Files are scoped to your account, used only to parse the package and answer your questions, and never visible to anyone else's account.
Which AI models does it use?
sangrep runs on Anthropic, OpenAI, or Google models. The rest doesn't change with the model: sangrep builds the context, checks every citation, and keeps the record no matter which one answers.
Do I need to clean up my files first?
No. Upload the bundle as it is. sangrep walks through it, shows you everything it found, and lets you decide how each embedded file should be treated.
What if something can't be parsed?
You'll know. Files that can't be parsed are flagged in the tree instead of silently skipped, so you can always see what's covered and what isn't.
Is my data used to train models?
No. Your documents are used to answer your questions and nothing else. sangrep doesn't train models on your data.
Can I plug it into my own tools?
That's the plan. sangrep will speak MCP, so the assistants and agents you already run can query your packages and get back cited, validated answers. The beta starts with the web app.
What does it cost?
Nothing while it's in private beta. Long-term pricing isn't settled yet, and people on the waitlist will hear about it first.

Be first through the door

One email when the beta opens. Nothing else.