Private AI. Verified work.
Private AI that does the work.
You approve it.
Recordrail answers questions over your company's documents and cites the passage every sentence came from. It runs on servers you control, and you choose which models may see which data.
- Self-hosted
- A citation on every sentence
- An audit log of every answer
How it works
From a folder of documents to an answer you can check.
Four steps, and a person at the end of them.
- 01
Bring your documents
Upload PDFs, Word files, spreadsheets, Markdown and images into collections. Each collection carries an egress class that decides which models may ever see its text.
- 02
Ask in plain language
Ask a question and, if you like, pick the model. Retrieval reads only the collections you are allowed to read.
- 03
Read the evidence
Every sentence carries a numbered citation that opens the source passage. When the documents do not support an answer, Recordrail says so instead of guessing.
- 04
You approve it
The question, the answer, its sources and the model choice are written to the audit log. What happens next is your call.
What it does
Five capabilities. One platform you run.
The same five pillars you see on the Home page once you sign in, with their status as it stands: Live is in the product today, Gated is granted per role, and Next is in build.
- 01Live
Governed AI
Data stays inside its egress class. Access follows the collection.
- Egress
- internal collections use only the internal gateway; the strictest class in scope wins
- Access
- retrieval is filtered to the collections you can read
- Audit
- every question and answer is recorded in the audit log
- 02Live
Evidence-grade answers
Every sentence cites a source. Unsupported questions are refused.
- Citations
- numbered anchors open the source passage
- Checks
- each claim is checked against its cited passage before it is shown
- 03Gated
Agentic workflows
Multi-step research runs with visible steps. Granted per role.
- Tools
- MCP access and service tokens are planned
- Experiments
- not yet available
- 04Live
Bring your own model
Pick the model per question. Internal data stays on internal models.
- Choice
- pick a model per question in Ask; the default follows the routing rule
- Control
- routing rules and fallbacks per task, set by an admin
- 05
Multimodal in and out
Scans and drawings become searchable. Answers point at the drawing.
- OCR
- scanned PDFs and images read by a self-hosted OCR model
- Regions
- citations will highlight the exact area of a page or drawing
- Status
- OCR stage in build; region answers follow
See the pillars with your own data.
Home shows each one against your collections, your benchmark runs and the models offered to you.
Sign in
Runs on your infrastructure
Your servers. Your models. Your rules.
Recordrail installs on hardware you own, including networks with no route to the internet. Which text may leave the box is a property of the data, enforced in code, never by convention.
Self-hosted, end to end
The API, the web app, Postgres, Qdrant, MinIO and Redis run on your own servers. There is no Recordrail cloud to send anything to.
Egress classes on every collection
Internal collections only ever reach the internal gateway. The strictest class in scope wins, and the model picker shows only what that scope allows.
Models you choose
Serve open models on your own hardware through vLLM or Ollama, or admit a cloud provider. Every call goes through one gateway, so a provider is configuration, not code.
Access that follows the collection
Admins invite people; there is no self-signup. Password plus a second factor, roles and permissions, and retrieval filtered to what each person may read.
An audit log, not a promise
Every question, answer and model choice is recorded with who asked and what was used, and admins can read it back.
- Postgres
- Qdrant
- MinIO
- Redis
Every answer shows its source
Cited, or refused. Nothing in between.
A fluent guess is the most expensive thing an assistant can give you. Recordrail is built so that it cannot.
- 1
Evidence before answers
The evidence check runs before any model is called. If the passages are not there, the answer is an honest refusal, never a fluent guess.
- 2
Citations that open the passage
Numbered anchors in the answer open the cited passage, with the document, its heading path and, for PDFs, the page.
- 3
Measured, not asserted
Retrieval and model changes ship with a benchmark run. A collection shows its trust score when one exists and says Not benchmarked when it does not.
What was decided about the second warehouse?
The documents in scope do not support an answer to this question.
UnsupportedNo model was called · recorded in the audit log
When the evidence is missing, this is the whole answer. The refusal is recorded like any other run, so you can see what was asked and what was not found.
Private AI. Verified work.
Private AI that does the work. You approve it.
Write to us for a walk-through on your own documents, or sign in if your organisation already runs Recordrail.