Conversational AI & RAG Agents

Chatbots and assistants that answer from your documents and show their sources.

What it is

A customer-support bot, an internal Q&A assistant or a product guide, built on content you already have: help-centre articles, PDFs, policies, a product manual. Retrieval-augmented generation (RAG) means the model looks up the relevant passages in your documents before it writes, so the answer comes from what you have published rather than from what the model remembers from training.

Your users type a question into a widget on your site or inside your app and get a direct answer with its sources listed underneath, each one linking to the document it came from. When the content does not cover the question, the bot says so and points to a human channel instead of guessing. Every session is logged, so you can read back what the bot said, which passages it used, and why.

What you get

  • Working bot with commented source code
  • Retrieval index and ingestion script
  • Evaluation suite and results
  • Deployment notes
  • Handover call

Every answer cites its source

How the work runs

The build steps once the quote is agreed. The brief, scoping call and quote are the same for every service.

  1. Content audit and ingestion

    Inventory the sources (help centre, PDFs, wiki, resolved tickets), decide what is in scope, and write the ingestion script that cleans, chunks and embeds them. It is re-runnable, so new content can be added later.

  2. Retrieval design and index

    Choose the chunking, the embedding model and the store (Chroma, pgvector or Pinecone), build the index, and check retrieval on 30 to 50 real questions before any answer is generated.

  3. Agent build with citations

    The answer prompt and format, the citation display, the refusal path for questions the content does not cover, and the embeddable widget for your site or app.

  4. Evaluation set and pass/fail receipts

    A held-out question set with known answers, graded deterministically. Every attempt in the run gets a pass/fail receipt, and the results ship with the bot.

  5. Launch and a week of tuning

    Deploy to Modal, Google Cloud Run or your own Docker host, read the first week of real sessions, and adjust retrieval and prompts from what they show.

Pricing

Starter

$1,000

Single-source RAG Q&A bot, embeddable web widget, citation display, one week of tuning after launch.

Enterprise

from $10,000

Multi-channel (web, Slack, email), custom retrieval pipeline, provenance ledger, uptime and response SLAs.

Prices are in USD and are starting points. The final quote follows a scoping call.

Where it fits

A good fit when

  • The answers already exist in writing: a help centre, manuals, policies, resolved tickets.
  • You need to know what the bot said and where it got it, for support quality or for a regulator.
  • Question volume has outgrown the people answering it.

Not a fit when

  • The answers are not written down yet. The bot can only cite what exists; if you are still deciding what to build first, start with AI Consulting & Proof-of-Concept.
  • You need a feature inside an existing app (summaries, drafting, classification) rather than a chat surface. That is AI Integration & APIs.
  • The task is a fixed transformation of text with no lookup, such as classifying tickets. That is Prompt Engineering & Fine-Tuning.

Questions

Where does my data live?

In your accounts. Documents are indexed into a vector store you control (Chroma or pgvector on your host, or your own Pinecone project), the bot runs under your API keys, and session logs are stored where you choose. GTDataWorks does not train models on your content and keeps no copy of it after handover.

Which model answers the questions?

Whichever fits the content and the budget: OpenAI or Anthropic models through the API, or an open model if the data cannot leave your network. You bring your own API keys, so usage is billed to you and you can switch providers later. The retrieval layer does not depend on which model writes the answer.

How is accuracy measured?

Against a held-out set of real questions with known answers, graded deterministically: did the answer cite the right passage, did it refuse when the content had no answer, did it stay within the source. You get the question set, the grading script and a pass/fail receipt for every attempt, so you can re-run the suite after any change to the content or the model.

What happens after handover?

You own the source, the index and the ingestion script, and the handover call walks through all three. Starter includes a week of tuning after launch. From there you can run the bot yourself, or add Monitoring & Analytics for dashboards and session replay over the logs it already writes.

Start with the documents you have

Send a link to your help centre or a sample of the documents, and five questions your users ask. That is enough for a scoping call and a quote.