CyberFreezeDev
Loading
An assistant that answers from your own files, cites each source and admits gaps.
The system finds the relevant passages in your documents and answers only from them, with links. Changed files need re-indexing, not retraining.
A general model like ChatGPT knows a lot about the world and nothing about your price list, your leave policy or the warranty terms you changed last month. A RAG chatbot fixes that without retraining the model.
Retrieval augmented generation works in two steps. First, every document is split into passages and indexed. When someone asks a question, the system searches that index and pulls the handful of passages most likely to hold the answer. Second, the model is given the question plus those passages and told to answer only from them, quoting where each fact came from.
The result is an assistant that a new hire, a customer or a field engineer can ask in plain language (English, Tamil, Hindi or a mix) and get an answer with a link to the exact page in the source document. When the documents change, you re index them. Nothing needs retraining.
A RAG chatbot answers questions; it does not act on your systems. If you need software that takes actions, read chatbot vs AI agent.
HR policies and SOPs in one place, so staff stop asking the same people.
HR policies, SOPs, process manuals and past circulars in one place. Staff stop asking the same five people the same questions.
Product and returns questions answered from help content, with handover to a person.
Answers product, shipping and returns questions from your published help content, and hands the chat to a person with the conversation attached when it cannot help. See WhatsApp CRM for the handover side.
Fault codes and torque values answered from the right manual revision, with page numbers.
Service engineers ask about a fault code or a torque value and get the passage from the right manual revision, with the page number.
Accounting policies, audit notes and circulars searchable by meaning.
Internal accounting policies, past audit notes and circulars, searchable by meaning rather than exact words. Useful for CA firms that keep years of client correspondence.
Most RAG prototypes work on ten clean PDFs and fail on a real shared drive. These are the parts we spend most of the time on.
Tables, columns and scans parsed with structure intact, using OCR where needed.
Tables, two column layouts, scanned pages and slide decks lose their meaning when flattened to plain text. We parse structure where it matters and run OCR on scans, using the pipeline from our document AI work.
Vector and keyword search combined and re-ranked, so exact codes are not missed.
Vector search finds passages with similar meaning. Keyword search finds exact codes, part numbers and section numbers that embeddings often miss. We run both and re rank the combined results.
Each passage carries its source's access rules, and search is filtered per user.
If a junior employee cannot open the board minutes in Drive, the assistant must not quote them. Each passage carries the access rules of its source and the search is filtered by the asking user.
Every change to chunking, search or prompts scored on 50 to 200 real questions.
We write 50 to 200 real questions with known answers together with your team. Every change to chunking, search or prompts is scored against it, so quality is measured rather than guessed.
When nothing relevant is found, it says so and offers a contact.
When retrieval finds nothing relevant, the assistant says the documents do not cover it and offers a contact. A confident wrong answer about a refund policy costs more than no answer.
Where knowledge lives, who owns it, how often it changes and who sees what.
List where the knowledge lives, who owns it, how often it changes and who may see what. Decide what is in scope for the first release.
Real questions from support logs and staff, each with its answer and source.
Collect real questions from support logs, emails and staff, with the correct answer and source for each.
Files pulled on a schedule, parsed, split with headings and page numbers, then indexed.
Connectors pull files on a schedule or on change, parse them, split them into passages with their headings and page numbers, embed and index them.
Chunk sizes, search weights and prompts adjusted until the question set passes.
Adjust chunk sizes, search weights and the answer prompt until the question set passes. Check citations by hand on a sample.
A small group rates every answer, and wrong ones become test questions.
Release to a small group with a thumbs up or down on every answer. Wrong answers become new test questions.
Opened to everyone in scope, with usage reports and a runbook for new sources.
Open it to everyone in scope, add usage and failure reports, and hand over a runbook for adding new sources.
Usually not. RAG is the right tool when the problem is knowledge: facts that live in documents and change over time. Fine-tuning is for teaching a model a format, style or narrow task. Many projects need only RAG.
With the business API terms of the main providers, inputs are not used for training by default, and we configure accounts accordingly. If documents must not leave your premises at all, we run the whole stack on your own servers. See private AI deployment.
We keep personal data out of the index where it is not needed, mask it where it is, apply access filters per user, and log who asked what. For UK and EU users the same controls support UK GDPR and GDPR obligations. Your legal team decides the lawful basis; we build the controls.
The connector notices the change, removes the old passages and indexes the new version. Answers cite the version they used, so you can see whether a reply came from the current policy.
Yes, but differently. For figures like stock levels or invoice totals we give the assistant a read-only query tool rather than indexing rows as text, so numbers come straight from the database.
Last updated