RAG & LLM Solutions
RAG development for your documents
RAG, or retrieval-augmented generation, is a way of building a chatbot that looks up your own documents before it answers, so replies come from your policies, manuals and records instead of the model's memory. Cognify builds RAG and LLM applications in India that cite their sources and admit when they do not know.
Scope
What is included
Document collection and cleaning
We gather PDFs, Word files, spreadsheets and web pages, remove duplicates and outdated versions, and agree which documents count as the source of truth.
Indexing for retrieval
Documents are split into passages, converted into searchable form and stored, so the system can find the few passages relevant to each question.
Answers with sources
Each reply names the document and passage it drew from, so a reader can check it, and the assistant declines when nothing relevant is found.
Access control
A staff member should not see another department's files through the chatbot. Permissions on documents carry through to what the assistant may retrieve.
Chat interfaces
A web chat on your site, an internal page for staff, or a WhatsApp number, depending on where your users already ask their questions.
Evaluation and updates
A test set of real questions with known answers, run after every change, plus an admin screen for adding or retiring documents without a developer.
Why it matters
Problems this solves
Repeated questions
Admissions desks, support teams and senior staff repeat answers that exist in a policy file. A grounded assistant answers first and escalates what it cannot.
Answers scattered across files and inboxes
Information sits in folders, old emails and individual memories. Searching by meaning, not by exact file name, finds what keyword search misses.
Chatbots that invent answers
A general-purpose model asked about your fee rules will guess confidently. Restricting it to retrieved passages, with citations, makes guesses visible and rarer.
Process
How we work
Gather
We collect the documents, ask who maintains each one, and mark which version is current. Weak source material is the commonest cause of weak answers.
Write test questions
You and we list real questions users ask, with the right answer and its source. This set becomes our measure of progress.
Build retrieval
We build and tune the search step first, since a model cannot answer well from the wrong passages.
Evaluate and tune
We run the test set, read failures together, and adjust chunking, prompts and rules until results are acceptable to you.
Deploy and monitor
After launch we review questions the assistant could not answer, which usually show missing documents rather than model faults.
Results
Example outcomes
An assistant that answers from your documents and shows where each answer came from.
A list of questions it could not answer, which doubles as a list of gaps in your documentation.
A test set you can re-run whenever documents or models change.
A way for your own staff to add and remove source documents.
Illustration
Example: an answer with its source
An illustration of a grounded answer. The question, policy name and page are invented for the example.
- QuestionIs there a late charge if the term fee is paid after the due date?
- AssistantYes. The fee policy says a late charge applies after the due date and describes how it is calculated.
- SourceFee Policy, section 4, page 6
- QuestionCan the principal waive it?
- AssistantI could not find this in the documents I have. I have passed your question to the accounts office.
It answers from the document where it can, and says so plainly where it cannot.
Technology we use
The retrieval layer and the model are separate parts, so either can be swapped as better options appear or as your data rules require.
Industries served
- Educational Institutes & SchoolsAnswers about admission rules, fee structures and regulations drawn from your prospectus and circulars.
- CA & FinanceAn internal assistant over firm checklists, past advice notes and compliance guidance, available to every team member.
- Custom EnterpriseStaff questions on standard operating procedures, product manuals and HR policies, answered in seconds with the source shown.
Pricing
How we price this
RAG projects vary with the volume and condition of your documents, the number of user groups, the interface you want and how strictly access must be controlled. Messy source files add preparation work, and we say so up front.
Language model usage is billed by the provider according to how much is asked, and hosting is billed by the host. We estimate both with you and show them separately from our build fee so there are no surprises.
FAQ
Questions, answered
What is the difference between RAG and fine-tuning a model?
RAG gives the model relevant passages at question time, so you update answers by updating documents. Fine-tuning changes the model itself and is slower and costlier to refresh. For company knowledge that changes, RAG is usually the better fit.
Will the chatbot make things up?
Any language model can make mistakes. We reduce the risk by limiting answers to retrieved passages, showing citations, setting the assistant to decline when evidence is weak, and testing against real questions before launch.
Where is our data stored and who can see it?
We agree the hosting location and provider with you, apply access controls, and explain what text is sent to a model provider. For sensitive material, we can discuss options that keep data in a private environment.
What file types can the assistant read?
Commonly PDF, Word, Excel, text and web pages. Scanned documents need text recognition first, which works well on clean scans and less well on poor ones. We test a sample of yours before committing.
Can it answer in Hindi as well as English?
Many current models handle Hindi and mixed Hindi-English questions reasonably, but quality varies with the model and the subject. We test with your real questions in both languages before promising a result.
How do we keep answers up to date?
Upload or remove a document in the admin screen and the index updates. We also recommend naming an owner for each source document, so outdated files are retired and the assistant stays accurate.
Let's talk
Tell us what you need
Share a few details and we will reply with how we would approach it. Prefer a chat? Message us on WhatsApp.
Chat on WhatsApp