RAG Chatbot on Custom Knowledge Base
A retrieval-augmented chatbot that finds the relevant passages by vector search and answers from them, citing the source document.
AI Consulting Services
AI consulting: a clear view of where AI will and will not help
A short, evidence-led assessment of your processes and data, ending in a ranked list of options, including the ones where AI is not the right answer.
Indicative starting price: $4,999 (USD). Final price depends on scope.
Technical notes
- Scores are agreed with you; they guide the order of work and are not a model output.
- "Data ready" covers volume, quality, access and whether personal data is involved (DPDP Act considerations).
- Where a rule or report is enough, we recommend that over a model.
What a pilot looks like
A pilot is narrow and has a decision at the end of each stage. We stop if the stage's exit criterion is not met.
Prepare data
Collect a slice of your real data, clean it and set aside examples to test against.
Exit criterionThe data is usable and a person can say what a correct answer looks like.Build and evaluate
Build the smallest approach that could work, and test it on the held-back examples.
Exit criterionMeasured results meet the success measure agreed at the start.Deploy with review
Run it on live work with a person checking the output, especially uncertain cases.
Exit criterionReviewers agree it saves effort and mistakes are caught.Monitor and learn
Track accuracy, collect corrected cases and decide whether to extend, retrain or stop.
DecisionExtend, adjust or retire it. Stopping is a valid outcome.
What the assessment report contains
- Candidate listEvery process considered, with the data behind it.
- ScorecardImpact, effort, data readiness and risk for each, as in the example above.
- Recommended first pilotOne use case, the data slice, and the measure that decides success.
- What not to doOptions we recommend against, with the reason.
- Privacy and security notesWhat personal data is involved and how the pilot would handle it.
- Next stepsWhat a pilot would take from you, and what a go/no-go would look like.
Problems this solves
- Leadership wants an AI initiative but nobody has defined the problem.
- A vendor proposal is hard to evaluate.
- An earlier pilot did not go to production.
- You are concerned about data privacy and accuracy.
What you receive
- A review of candidate processes and the data behind them
- A ranked shortlist with effort, risk and how success would be measured
- A recommended first pilot, scoped on a slice of your real data
- Notes on privacy, security and the DPDP Act where personal data is involved
How the work runs
Each phase has an input from you and a written output from us. Building the pilot itself is a separate engagement.
| Phase | What happens | You provide | You receive |
|---|---|---|---|
| Step 1Review | Listen to where time and money go. List candidate processes and check what data exists behind each. | A few hours with process owners; examples of the data. | A candidate list with notes on data readiness. |
| Step 2Score and shortlist | Score each candidate on impact, effort, data readiness and risk, including privacy. | Input on business priorities. | A ranked shortlist, including options where AI is not the answer. |
| Step 3Recommend | Choose a first pilot and define the measure that will decide whether it worked. | Agreement on the success measure. | A written recommendation. |
| Step 4Plan the pilot | Scope a narrow pilot on a slice of your data with a person reviewing the output. | Confirmation of the data slice and reviewer. | A pilot plan with go/no-go criteria. |
Related work
Anonymised client projects. Client names are withheld because of confidentiality; the project page explains each one.
From the blog
Practical articles from the engineers on this topic.
The Future of AI in Business: Trends to Watch in 2026
Forget the headline-grabbing demos. The changes that are reaching ordinary businesses this year are narrower, cheaper and less glamorous than the coverage suggests. Here is what looks durable, …
FAQ
Frequently asked questions
Can you extract data from invoices and PDFs?
Yes. We have built OCR-based extraction pipelines that validate fields against business rules, send mismatches to a person for review and keep an audit trail.
Do we need to be 'AI ready' before we start?
No. The assessment itself shows what is ready and what is not. We start from one measurable problem and a slice of your data.
How do you test a RAG chatbot before launch?
With a fixed set of real questions, including ones that have no answer in your documents. We score retrieval and answers separately and agree a pass threshold before launch.
Do you recommend AI for everything?
No. Where a rule, a report or a simple automation does the job, we say so. AI is used where it earns its place.
What does a first pilot look like?
A narrow use case running on a slice of your real data, with a person reviewing the output, so you can judge accuracy before committing further.
How is the work priced?
The 'starting at' figure is indicative. After a short discovery conversation we share a written scope and an estimate based on what the project actually involves.
Who owns the code and the data?
This is agreed in writing before work starts. Our default is that you own what we build for you, and that your data stays in systems you control.
Told to 'do something with AI'?
Tell us which processes cost the most time. We will give you an honest view of where AI will and will not help.