Answers you can check.
We build private AI assistants over your company's documents. Every answer points to the page it came from, and it all runs on infrastructure you control.
The retrieval engine is open source under the MIT license, and uses local models by default.
QuestionWhat was 3M's capital expenditure in fiscal 2018?
3M spent $1,577 million on capital expenditure in 2018, reported in its cash flow statement as purchases of property, plant and equipment.
| Line item | 2018 | 2017 | 2016 |
|---|---|---|---|
| Net cash provided by operating activities | 6,439 | 6,240 | 6,662 |
| Purchases of property, plant and equipment (PP&E) | (1,577) | (1,373) | (1,420) |
| Proceeds from sale of PP&E and other assets | 262 | 49 | 58 |
| Acquisitions, net of cash acquired | 13 | (2,023) | (16) |
A research assistant for your own documents
You bring the contracts, policies, filings, manuals or support docs your team keeps searching through. We set up an assistant they can question in plain language, tuned and tested on your material.
- Cited answers
- Every answer names the file and page it used, so anyone can verify it in a click.
- Runs where your data lives
- Your servers or private cloud, with open models by default. Nothing leaves your network unless you choose a hosted model.
- Connected to your sources
- File shares, SharePoint, Confluence, databases. We build the connectors your documents need.
Built per engagement - Access that matches your org
- People only get answers from documents they are allowed to open, wired to the sign-in you already use.
Built per engagement - Tested before launch
- You get an evaluation report on questions written by your own team, including the ones it gets wrong.
What happens to a question
A single vector search misses exact figures, names and follow-ups. Cortex runs each question through a sequence of checks before anything is written.
Understand the question
Follow-ups like "and the year before?" are rewritten into a complete question first.
Search two ways
Keyword search finds exact terms and figures. Semantic search finds passages that mean the same thing in other words.
Follow related names
An entity graph pulls in passages about the same companies, people and products.
Rerank
A cross-encoder reads every candidate against the question and keeps the strongest.
Check relevance
The model grades each passage and drops the ones that don't answer the question.
Answer with sources
The answer cites each passage it used, down to the page.
Contextual retrieval, HyDE and RAG-Fusion are available for harder collections. Every step can be switched on or off, and we measure which ones actually help on your documents before turning them on.
Measured, not claimed
We ran Cortex on FinanceBench: 150 questions written by financial analysts about 84 real annual and quarterly reports. Everything ran locally on a laptop with an 8B open model.
Two inexpensive changes did most of the work: sending each question to the filing it names, and reranking the passages before answering. When the evidence isn't found, Cortex says so instead of guessing. Questions that need arithmetic are still the weak spot, and that's what we're improving next.
Answers were graded by a local model with strict rules, where a refusal counts as wrong, not by people. For context, the 2023 FinanceBench study found GPT-4 Turbo with a standard retrieval setup got 81% wrong or refused; the setups differ, so it isn't a head-to-head. Full results and code.
| Plain RAG | Cortex | |
|---|---|---|
| Evidence page in top 5 | 25% | 39% |
| Right filing in top 5 | 78% | 93% |
| Answers correct | 36% | 45% |
| Said "not found" instead | 55% | 43% |
How a pilot works
A 30-minute call
Tell us what your team keeps searching for. We'll tell you honestly whether retrieval can fix it.
A pilot on your documents
We index a sample, write test questions with you and share the scores, misses included. We aim to finish in about two weeks.
Deployment
We install it on your infrastructure, connect your sources and hand over the evaluation suite so you can keep measuring.
Read the code before you call us
The engine is open source. Clone it, run it on your laptop against your own files, and see whether the answers hold up. If they do, we can build the version your organisation needs.
Or email saiakhil066@gmail.com.