06 ·Agnidoot RAG· Enterprise
Not a search box over a folder. An advanced retrieval system that reads your documents, pulls out the entities, processes and figures inside them, connects them to the records already in your Odoo — and makes the whole picture available to your ERP and your AI agents at once.
In plain language
Skip the technology for a moment. Here is the difference it makes on an ordinary Tuesday.
It's the difference between having all the files — and having someone who has read them all.
Every business already owns the answers. They're in contracts, emails, quotations and QC reports. The problem was never storage. It's that finding anything means knowing where to look, and remembering it exists at all.
The pattern in all four: the answer already existed inside your business. It just wasn't reachable in the moment someone needed it — and the cost of that is paid in slow decisions, repeated work and mistakes nobody sees coming.
The idea
Your best employee is the one who's been there fifteen years — who remembers which supplier was late last winter, what that contract actually promised, and why the customer in Pune gets different payment terms. That knowledge lives in their head, in documents nobody can find, and it walks out at six o'clock.
Intelligent context is that memory, built from your own paperwork — and unlike a person, it reads every document, forgets nothing, and cites where each fact came from.
Every party mentioned across every document, resolved to one entity even when four spellings exist — and linked to the partner record in your Odoo.
How work actually flows, and the commitments buried in the text — "valid 30 days" becomes a number, and your rules stay separate from your counterparties'.
Amounts, quantities, deadlines and identifiers extracted as structured values — checkable against the ledger rather than buried in a paragraph.
The pipeline
Most "chat with your documents" tools do two of these: chunk and embed. The five in the middle are why the answer knows what a vendor is.
Apache Tika with OCR, plus local fallbacks so a single service being down doesn't stop the batch. Scanned PDFs are detected and sent through image extraction with OCR language headers, rather than returning an empty page.
Twenty-six formats, verified from the code:
.pdf.doc.docx.txt.md
.rtf.odt.ods.odp.ppt
.pptx.xls.xlsx.html.htm
.xml.csv.json
.png.jpg.jpeg
.tiff.tif.bmp
.webp.gif
Green ones are images, read by OCR. A corrupt or empty file is skipped and named in the response; the rest of the batch processes normally, and a server restart mid-job doesn't strand it.
Every document gets a type, a domain and a confidence score the moment it's processed — no manual tagging needed to get started. A document the system couldn't classify shows an amber "needs review" badge rather than a misleading green "processed", and a correction you make by hand survives a workspace-wide re-check untouched.
This is the stage that separates intelligent context from a search index. Each document is read for meaning, not just text:
"Apex Metering Systems Private Limited", "Apex Metering Systems" and "Apex Metering" are one vendor. Without this stage, a search for one finds a third of the truth.
Documents are split into passages, embedded by a model running on the AI-ready box in your building, and written to two indexes at once — a keyword index and a vector index, both isolated per tenant.
Retrieval queries both and fuses the results, so a search finds the exact phrase you typed and the passage that means the same thing in different words.
Exact terms, part codes, invoice numbers, names — the things a vector search is famously bad at.
Passages about late delivery, when what you typed was "supplier reliability problems".
And when half of it is unavailable, it says so. A degraded search raises an error rather than returning keyword-only results dressed up as a complete answer. "No results" and "I couldn't look properly" mean very different things to somebody making a decision.
Where the context shows up
The point of building context is not to have a second place to search. It's that everything else in the business gets smarter at the same time.
The vendor you're looking at carries what the documents say about them — terms, history, the complaint from March — instead of living in a folder somebody else owns.
Ask about overdue invoices for customers with quality complaints and get one answer — ERP records and document context together, scoped to what you're allowed to see.
Group documents by project, client or department, and share a workspace with a person, a team or a group — each with its own permission level.
Business documents don't live in one place. They're on laptops, in cloud drives, in object storage and on a server somebody set up years ago.
There is no desktop sync app yet. Device registration exists; a proven path for a file moving from a laptop into the corpus does not. Today the routes in are the connectors above and direct upload — so if your documents live on individual machines, the answer is a synced cloud drive rather than an agent on each laptop. Treat desktop ingestion as unbuilt until someone demonstrates it to you.
There's no file-server connector either — SMB and NFS shares aren't reachable today. And Gmail ingestion was removed in August 2026 because it produced fabricated data; we'd rather ship no connector than a lying one.
Private by construction
Turning a document into something searchable means turning it into numbers. Where that happens decides whether your contracts leave the building.
One customer's document surfacing in another's search isn't a bug to fix next release, so isolation sits at the storage layer rather than in a query somebody must remember to constrain.
The embedding model runs locally, so document text never leaves the deployment — and we hold that as a contract term, not a checkbox, because repointing it would change the answer silently.
An estate is only as safe as the weakest permission on it.
The details
No — and that was a deliberate decision, not an oversight. Media transcription was explicitly taken out of scope and the code removed. The pipeline handles 26 document and image formats, including scans and photographs through OCR.
If meeting recordings or site videos matter to your use case, tell us during the pilot and we'll be straight about whether it's something we'd build rather than implying it already works.
We separate these deliberately:
No penetration test or compliance certification has been performed on any part of the platform.
Yes. Point both the embedding model and the language model at local hardware and no document text or prompt content leaves your network, including air-gapped deployments. This is the configuration regulated customers should ask for by name.
Every plan includes
No obligation, no card, no lock-in. And what comes with your plan isn't a discount — it's what it takes to make an ERP actually land.
Not months. Your implementation is underway in your first week.
A full year included with every plan — not a paid add-on.
We come to you and train your team in the room, on your data.
A hundred hours of our engineering, built into every plan.
◆ No card to start · ◆ No obligation · ◆ Cancel any time in the three months · ◆ Your data stays yours
Questions
The one that spans your ERP and a folder of documents, and currently costs somebody an afternoon. We'll answer it on your data — and show you the sentence each fact came from.
Thirty minutes, your paperwork, your questions. We will show you what it does — and, just as importantly, what it refuses to do.