ATPL

AI · Build with AI

Answers drawn from your own documents, with sources.

For companies whose answers live in manuals, policies, contracts and old emails: staff and customers ask in plain words and get the answer with the page it came from.

Describe the problemAll AI capabilities

Buried docs

Answers buried in folders

Answers drawn from your own documents, with the source page shown.

Describe the problem

  1. Search

    New staff ask the same questions for months because the answer is in a manual nobody opens.

  2. Wrong

    Customers wait a day for a reply that is already written in your policy document.

  3. Access

    A public chatbot gave confident answers that were wrong, and you cannot afford that with your own material.

Your data

Your documents are indexed inside infrastructure you control and are never uploaded to a model provider as a whole; each question sends only the few passages retrieved for it. Provider terms exclude training on your data. Questions and answers are logged so quality can be checked, and the log is yours to keep or delete.

Search or assistant

Where it fits, where it doesn’t

  • Fits when
  • Questions whose answers exist in writing: manuals, policies, contracts, SOPs, past tickets.
  • Teams that answer the same questions repeatedly and can tell a right answer from a wrong one.
  • Document sets that change slowly enough for a scheduled re-index to keep up.
  • Doesn’t fit when
  • Questions that need judgement or live data the documents do not hold, such as today’s stock.
  • Document sets that are themselves wrong or contradictory; the assistant will repeat the contradiction faithfully.
  • Fewer than a few dozen pages. At that size a well-organised FAQ page is cheaper and just as fast.

Answers with citations

An assistant on your documents

  • Answers with citations

    Every answer shows the document and section it came from, so a person can check it in one click.

  • Permission scoping

    The assistant only retrieves what the asking user is allowed to read; HR files do not surface for a sales query.

  • A refusal path

    When the documents do not contain the answer, the assistant says so and routes the question to a named person instead of guessing.

  • Question log and evaluation set

    Unanswered and badly answered questions are reviewed weekly and added to a test set of real questions with agreed answers.

Primary deliverable

Ships with the engagement

Live brief

01 · Lead outcome

A document index you own

Your PDFs, Word files, wiki pages and tickets are split, embedded and stored in a search index inside your infrastructure, refreshed when files change.

  • 01Answers with citations
  • 02Permission scoping
  • 03A refusal path
  • 04Question log and evaluation set

Ask your corpus

A question becomes a cited answer

The brief is sources and confidence, not a confident guess. If the corpus does not have it, the system should say so.

What a client typed

Our dealers keep calling the office to ask about warranty terms, spare part numbers and installation steps, and all of it is in the product manuals and the circulars we have sent over the years as PDFs on a shared drive.

Brief ready

What it becomes

Problem
Dealer questions answered by phone from information that already exists in manuals and circulars.
Scale
A dealer network asking daily; a shared drive of PDFs built up over years.
Systems
Product manuals and circulars on a shared drive, the dealer portal or WhatsApp.
First step
An assistant for office staff first, answering with the manual page shown; dealers get it once accuracy is proven.

How it ships

Index, cite, refuse

  1. rag/01

    Understand

    We take your documents as they are and a list of fifty real questions staff or customers have actually asked.

  2. rag/02

    Design

    We decide chunking, the index, who may see what, and the exact wording for "I don’t know" before any model is chosen.

  3. rag/03

    Build

    Retrieval is tuned first, because a model cannot answer from a page it was never shown; then the answer step is added and scored against the question set.

  4. rag/04

    Stay

    The index re-syncs on a schedule, answer quality is sampled monthly, and new document types are added as they appear.

How we work

This website’s own brief assistant on the Contact page answers only from the site’s text, built on the same retrieve-and-cite method; ATPL will confirm which client deployments may be named.

    Chosen for the job

    Python · Node.js · Next.js

    Decide the shape

    Before you start

    Plain answers on scope, data and what a first proof usually takes.

    RAG, retrieval-augmented generation, is a method where a search step first finds the relevant passages in your own documents and a language model then writes the answer from those passages only. It is how ATPL builds knowledge assistants that cite their sources instead of answering from general training.

    Accuracy depends mostly on retrieval, which is why ATPL tunes the search step first and scores it against fifty real questions before adding the answering model. A well-built assistant answers correctly when the answer is in the documents, cites the page, and says "not in the documents" when it is not.

    Yes, with an OCR step before indexing for scanned pages and with embedding models that handle Hindi, Gujarati and other Indian languages. ATPL checks retrieval quality per language on your real files rather than assuming it, because scan quality and language mix differ from company to company.

    It can be, if permissions are enforced at retrieval time. ATPL builds knowledge assistants so that each user only receives passages from documents they are already allowed to open, the index stays in your infrastructure, and nothing is used to train anyone’s model.

    Continue to