A custom chatbot is a retrieval system with a conversational interface. Understanding that shapes the whole project: most of the effort goes into what it knows and how it finds things, not into how it talks.

Here is the process we follow.

Stage 1: Audit the content

Before anything technical, inventory what the bot should know and assess its state.

Collect the sources: website pages, product data, policy documents, help articles, PDFs, past support tickets, internal documentation.

Then assess honestly:

  • What is current, and what is superseded?
  • Where do documents contradict each other?
  • What is in a format that can be indexed, and what is trapped in a scanned image?
  • What should definitely not be indexed — drafts, internal pricing, anything confidential?

This stage is unglamorous and determines everything downstream. A bot indexed on contradictory content gives contradictory answers, and users lose trust permanently after two or three of those.

The audit is also independently valuable. Clients routinely discover their published policies disagree with each other.

Stage 2: Build the question set

Before building the bot, collect real questions with known correct answers.

Sources: support tickets, sales call notes, the site search log, Search Console queries, and simply asking the people who answer customer questions all day.

Include deliberately:

  • The common questions, in proportion to real volume
  • Awkward phrasings and misspellings
  • Questions the content cannot answer, where the correct response is "I don't know, let me connect you"
  • Questions about things that do not exist — a product you do not sell
  • Questions that should escalate: complaints, anything emotional

Thirty to a hundred questions is enough. This set is what makes every later decision measurable rather than a matter of opinion.

Stage 3: Build and tune retrieval

The engineering core.

Chunking. Split documents in a way that suits their structure. Policies split by clause, help articles by section, tickets as single units. A single strategy applied to everything is the most common reason retrieval underperforms.

Embedding and indexing. Convert chunks to vectors and store them, with metadata attached — document type, date, access level, source.

Hybrid search. Combine semantic and keyword search. Semantic handles "can I send this back"; keyword handles "SKU-4471". Neither alone is sufficient on real business content.

Re-ranking. Retrieve more candidates than needed, then re-rank for relevance before passing the top few to the model. This produces a larger quality improvement than almost any other single change.

Tune against the question set. Measure how often the correct passage is retrieved at all. If retrieval is wrong, no model will save the answer.

Stage 4: Persona and guardrails

Now the conversational layer, which is quicker than people expect.

Persona. Tone, response length, formality. Keep responses short — a wall of text in a chat window gets skimmed and distrusted.

Grounding instruction. Answer only from the provided context; if it is not there, say so and offer a human. State what to do instead of answering, not just what not to do.

Refusal as success. Make clear in both the prompt and your evaluation that declining is a correct outcome. If your scoring rewards answering, you will optimise toward confident invention.

Escalation triggers. Complaint language, explicit requests for a person, repeated failure, high-value accounts, anything outside scope.

Hard limits. Things it must never do: quote prices it generated rather than retrieved, promise outcomes, make commitments on your behalf.

Stage 5: Test properly

Run the question set. Score on:

  • Correct answers given
  • Appropriate refusals on unanswerable questions
  • False confidence rate — confident answers to questions it should have declined. The most important safety metric.
  • Appropriate escalations

Iterate on retrieval and prompts until the numbers hold up. Then have people outside the project use it, because they will ask things nobody on the team thought of.

Stage 6: Deploy to one channel

Launch on web chat first, not everywhere.

Make the human route obvious. One click, visible at all times. Hiding it to improve deflection metrics is the most counterproductive decision available — visitors who suspect they are trapped fight the bot from the first message.

Design the handoff. The human receives the full conversation, what the bot understood, what it attempted, and the customer's context already looked up. Nobody should repeat themselves.

Capture leads naturally. Mid-conversation when the visitor is engaged, not as an upfront gate.

Stage 7: The weekly loop

This is what turns a decent launch into a good system.

Every week:

  1. Read the unanswered-question log. Each entry is a content gap.
  2. Read a sample of handled conversations. Not just the flagged ones.
  3. Add content to close the gaps. Most problems are fixed by writing something, not by changing the bot.
  4. Add new failures to the question set, so the same mistake cannot recur.
  5. Re-index and re-run the evaluation.

The unanswered-question log is the most valuable output of the entire system. It tells you exactly what your customers want to know and cannot find — which improves your website and your search visibility as well as the bot. Several clients have found their best-performing content this way.

Then expand

Once web chat is reliable:

  • Add WhatsApp, sharing the same knowledge base. One source of truth, not two bots that drift apart.
  • Add Messenger and Instagram if those are channels your customers use.
  • Connect live data — order status, availability, account details — which is what moves a bot from informative to genuinely useful.

Timeline

A focused support or sales bot on reasonably organised content: two to four weeks.

Larger builds spanning multiple channels, languages and system integrations: six to ten weeks.

The variable is almost always content state and integration complexity, not the conversational part.

The thing that determines success

A chatbot is a mirror of your documentation. If your content is current, clear and comprehensive, the bot will be good. If it is contradictory and out of date, the bot will be an articulate, confident mess.

Most of what makes a chatbot project succeed happens before anyone writes a prompt.

If you want to know whether your content is in good enough shape to build on, send us what you have — the audit is usually a short piece of work and useful regardless.