It answers from your documents
Retrieval first, generation second. The model summarises what was found rather than recalling what it was trained on.
A chatbot that answers from your content, not from its imagination.
Most business chatbots fail for one of two reasons: they answer from a decision tree nobody maintained, or they answer from a general model that cheerfully invents your refund policy. I build the third kind – one that reads your own documents and says so when it does not know.
A language model on its own knows nothing about your business. Ask it about your delivery timelines and it will produce something plausible and wrong, because producing plausible text is the whole of what it does. That failure mode is not a bug to be prompted away; it is the technology working as designed.
The fix is retrieval. Your documents – product pages, policies, price lists, past support replies – are split, indexed and searched at the moment of each question. The model is handed the relevant passages and asked to answer from those alone. It is the difference between asking someone to recall your policy and handing them the policy to read aloud.
That single change moves the project from a novelty to something you can put in front of customers. It also makes the failure mode safe: when nothing relevant is found, the honest answer is “I do not have that” and a route to a person, rather than a confident fabrication that you find out about from an angry email.
What it does not do is replace your support team. It removes the thirty repeated questions a day that never needed a human, and it hands over everything else with the conversation attached. A chatbot sold as a replacement for staff is being oversold.
A widget on your site that answers from your own pages, documents and policies, in your tone rather than a vendor’s.
The questions that arrive at 11pm get answered at 11pm.
Your PDFs, policy pages and product data indexed so answers are retrieved from them and cite where they came from.
You can check every answer against the source it used.
Conversations that collect what you actually need – requirement, budget range, timeline – before a human is involved.
Your inbox fills with qualified enquiries rather than “hi”.
The same indexed content serving a web widget and a WhatsApp number, so both say the same thing.
One place to update when a policy changes.
Clear escalation rules, with the whole conversation passed to whoever picks it up.
Nobody has to ask the customer to start again.
Auditing an existing bot, finding where it is answering ungrounded, and putting retrieval under it.
A bot you can stop apologising for.
Six things that follow from grounding answers in your own content.
Retrieval first, generation second. The model summarises what was found rather than recalling what it was trained on.
No relevant passage means no answer and a route to a person – the single most important behaviour to get right.
The repeated questions get answered at night and at weekends, when they would otherwise wait.
Escalation rules you set, with the full conversation attached so nothing is repeated.
Logged questions are a content brief: the gaps in your site, written by your customers.
Wrong answer? Fix the source document and reindex. You are editing content, not prompts.
And who is better served by something simpler.
Where support time goes on questions your own site already answers badly.
Sizing, compatibility, stock and delivery questions that arrive before every order.
Policies, scope documents and price lists that customers will not read but will ask about.
Timings, preparation instructions and what to bring – asked constantly, rarely urgent.
Where enquiries arrive on a personal number and nothing is logged or qualified.
With little content to retrieve from, a chatbot has nothing to ground on. A clear FAQ page is the honest answer.
Indian customers ask before they buy, and they ask in the channel they already have open. For most businesses here that is WhatsApp rather than a web form, and the enquiry arrives as a voice note or a one-line message at an hour nobody is working. The question is rarely difficult – it is the same question as yesterday – but it still costs a person twenty minutes to find and answer.
Language is the part people underestimate. Customers write in English, in Hindi, in Kannada, and frequently in two at once, transliterated. A bot trained on a rigid script fails on the first message. A retrieval-based one copes far better, because the matching happens on meaning rather than on exact phrasing – though it is worth testing with your own customers’ real messages before anyone promises anything.
On cost: these systems are priced per conversation and per token, and those numbers are small until they are not. I model the running cost against your actual message volume before you commit, so the bill in month three is one you already saw in month zero.
A WhatsApp number and a web widget answering from one source.
Tested against how your customers actually write, not an idealised script.
Per-conversation economics shown before you commit, not discovered later.
Handover routed to people who are awake, by rule rather than by hope.
Serving Chennai, Bangalore, India, Canada, UK, USA, Dubai, Singapore.
The pattern is the same: high question volume, low question variety.
Typically needs: Pre-purchase questions on sizing, compatibility, delivery and returns.
How I help: Catalogue and policy pages indexed, with order-status handover to a person.
Typically needs: Timings, preparation, documents to bring, what is covered.
How I help: Practice documents indexed, with anything clinical escalated immediately and by rule.
Typically needs: Course content, fees, schedules, eligibility, placement questions.
How I help: Prospectus and course pages indexed, enquiries qualified before a counsellor calls.
Typically needs: Availability, configuration, location, documentation.
How I help: Listing data and FAQs indexed, with site-visit requests captured properly.
Typically needs: Scope, process, what a given engagement includes.
How I help: Scope documents and past proposals indexed, with budget qualified before a call.
Typically needs: Setup, integration and troubleshooting questions that are in the docs.
How I help: Documentation indexed, with a ticket raised when retrieval finds nothing.
Six steps, with a real decision point at the end of the first.
I read your last few hundred real enquiries and work out what proportion a grounded bot could answer. We look at what content exists to retrieve from, and what does not.
You get: A written scope, a fixed quote, and a straight answer on whether this is worth building.
Why it matters: The projects that fail were the ones nobody checked had enough content behind them.
The content set is decided: which pages, documents and policies are in scope, who owns each, and how often they change. Escalation rules are written down.
You get: A documented source list and a handover policy, agreed before anything is built.
Why it matters: A bot is only ever as current as the documents under it. Ownership decided now prevents staleness later.
The conversation is designed: opening message, tone, what it asks for, what it never attempts, and exactly how it says it does not know.
You get: A written conversation design, including the refusal and handover wording.
Why it matters: How a bot declines is more important to your reputation than how it answers.
The retrieval pipeline and the integration are built: documents chunked and indexed, the model wired up, the widget or WhatsApp number connected, logging in place.
You get: A working chatbot on a staging URL, with conversation logs you can read.
Why it matters: Seeing real retrieval on your own content is the only way to judge it.
It is tested against your real past questions, including the awkward ones, and tuned where retrieval misses. Cost per conversation is measured.
You get: A test report showing answered, escalated and missed, with the running cost per conversation.
Why it matters: Testing on invented questions proves nothing. Your own backlog is the honest benchmark.
It goes live with monitoring, and I stay on the logs for the first weeks while the real questions arrive.
You get: A live chatbot, a handover that works, and a monthly review of what it could not answer.
Why it matters: The first month of logs is the most useful content brief you will get all year.
Chosen per project, and explained before anything is committed to.
The parts that are not optional.
A chatbot is not an SEO tactic. It sits behind a click, its answers are not indexable, and a widget that blocks content or shifts layout will cost you more in Core Web Vitals than it ever returns in engagement.
What it does give you is the best content brief available: a log of what real people asked and what your site failed to answer. Turning the top twenty of those into proper pages is the part that earns search traffic – the bot just found them for you.
No one can promise that ChatGPT, Gemini, Perplexity or Google’s AI Overviews will cite your site, and a chatbot on your pages has no bearing on whether they do. What is within your control is whether your content is clear, structured and attributable – which helps a retrieval system of any kind, including your own.
Four things I will do that a chatbot vendor will not.
If your last two hundred enquiries are all different, retrieval has nothing to stand on and a chatbot will frustrate people. That conversation happens before the quote, not after the invoice.
Twelve years of WordPress and WooCommerce means the bot can read real product, order and policy data rather than a copy that drifts out of date.
Per-conversation pricing is where these projects surprise people. You see the arithmetic against your real volume before you commit to anything.
The index, the logs and the integration code are yours. If you want to move it in-house or to someone else, nothing here is designed to stop you.
Tell me what you need. I reply within one working day, and the first conversation costs nothing.