Build a niche answer assistant on video transcripts

Every specialised field has a decade of recorded talks nobody can search. Conference keynotes, training webinars, certification courses, the association’s monthly session — four hundred hours of the field’s actual working knowledge, indexed by title only. Someone with a question has to already know which talk answered it.

An assistant that answers from those transcripts is a genuinely useful thing to build, and the corpus costs about two dollars. This playbook covers what it takes to make it trustworthy enough to sell, which is most of the work, and the part where a language model will confidently invent an answer if you let it.

What you are actually selling

Institutional memory, made askable. The client owns years of recorded expertise that currently produces value once, on the day it is recorded. You are turning it into something a member can query at eleven at night and get a cited answer from.

The word that carries the sale is cited. A generic chatbot with the field’s name on it is worth nothing and everyone has already seen one. An assistant that answers “the 2023 session on this said X” and links to the timestamp in the video is worth a build fee, because the citation is what makes the answer usable in the member’s own work.

The recurring fee is not for hosting, which is cheap. It is for the corpus staying current — new sessions ingested as they publish — and for someone answering when it gets an answer wrong.

The numbers

The data is the cheapest line by a wide margin. Price the build on the integration and the evaluation work, because that is where the weeks go.

Ninety hours is realistic for the first one and about thirty-five for the third, once you have a pipeline you trust. The build fee is close to break-even on your time by design: it buys the relationship, and the recurring line is the business. Six clients at $800 is $4,800 a month against roughly $600 of cost and a few hours of maintenance.

Refreshing the corpus is trivially cheap — fifty new talks a quarter is twenty-five cents — so never meter it. Charging for ingestion makes the client ration the thing that keeps the product useful.

How it works

1. Confirm the client owns or is licensed for the content before anything else. This is the first question, not the last. See the Limitations below; it is what makes the difference between a product and a liability.
2. Assemble the corpus from a defined search. Search the field’s terms, take the talks from the channels the client recognises, and keep the video ID, title, channel, publication date and duration alongside the text. Those fields are what every citation is built from later.
3. Chunk on meaning, and carry timestamps through. Chunks of roughly two hundred words with overlap, each one stamped with its start time. The timestamp is what turns a citation into a link that lands on the moment, and retrofitting it later means re-chunking everything.
4. Store the source metadata with every chunk. A retrieved chunk that cannot say which talk it came from is unusable, and this is the most common thing to leave until it is expensive to add.
5. Make refusal the default. The system prompt must say that if the retrieved passages do not answer the question, the answer is that the corpus does not cover it. This costs you some coverage and buys the entire credibility of the product.
6. Build the evaluation set before the interface. Fifty real questions from the client’s members, with an answer each expert agrees on. Run it on every change. Without this you cannot tell an improvement from a regression, and you will ship both.
7. Show the sources beside every answer. Talk title, speaker, date, and a link to the timestamp. Members will click them, which is the point: the assistant is a way into the content rather than a replacement for it.
8. Ingest new talks on a schedule and say when the corpus was last refreshed. A visible “current to” date is a small feature that does a lot for trust.

Put a one-click “this answer is wrong” control on every response and read what comes back weekly. It is your best source of evaluation questions and the clearest evidence of value at renewal.

Where it gets hard

The model will invent things. Confidently, fluently, in the client’s own vocabulary. Retrieval reduces it and does not remove it. In a professional field a plausible wrong answer is worse than no product, so the refusal behaviour and the evaluation set are not polish — they are the deliverable, and most of the ninety hours goes into them.

Transcripts of talks are messy input. Speakers correct themselves, slides carry half the meaning, and a question from the floor is often inaudible. Some talks are not usable as text at all and you will not know which until you read them.

The client will ask for it to answer things the corpus does not contain. Their handbook, their pricing, their member directory. Each of those is a separate integration and should be a separate quote, or the project has no edge and never ships.

Selling to associations is slow. Committees, budget cycles, an annual meeting where it gets decided. Build the demo on their real content before the first meeting — the ten minutes where they ask it something only an insider would ask and it answers with a citation is the entire sale, and no proposal reproduces it.

Limitations

The corpus is someone’s spoken words, and a recorded talk is the speaker’s work. Building an assistant over a client’s own content, or over content they are licensed to use, is one thing. Building one over a third party’s talks and selling access to it is republication, and the fact that the output is paraphrased does not change that. Establish the rights position first, in writing, and decline the ones where it is unclear.

Retrieval quality is bounded by the corpus. If the field’s answer to a question was never said on video, no amount of tuning produces it, and an assistant that fills that gap from the model’s general knowledge is worse than one that says nothing. Grounding must be strict.

Transcripts do not capture slides, diagrams, code on screen or anything demonstrated visually. In technical fields this is a large fraction of the content, and any question whose answer was on a slide will be answered badly or not at all. Say so to the client before they discover it.

Automatic captions carry errors, and they cluster exactly where you least want them: proper nouns, drug names, product names, acronyms. Those are the terms members search for. Budget for a glossary of the field’s fifty most important terms and their common mistranscriptions.

Nothing here is a substitute for professional advice in a regulated field. If the client is in medicine, law or finance, the disclaimer and the refusal behaviour need to be reviewed by someone qualified, and that review is a line item.

FAQ

How many talks make a useful corpus?

Two hundred is enough to be useful in a narrow field; four hundred to a thousand is comfortable. Breadth matters more than depth — a hundred talks spanning the field beats four hundred from one annual conference, because the failure mode members notice is the question that returns nothing.

Can the client not do this themselves now that the tools are easy?

Some will, and they will get a demo that works on ten questions and fails on the eleventh. What you are selling is the part that does not demo: the evaluation set, the refusal behaviour, the citation plumbing and someone to call when an answer is wrong. Say that plainly rather than competing on the demo.

Should I charge for the build at all if the margin is thin?

Yes. A free build is treated as a free build, and it takes the client’s commitment away at the exact moment you need them to supply content, feedback and an internal champion. Price it near your cost and let the recurring fee be the business.

What happens when the client cancels?

Agree at the start what they get: their corpus and their evaluation set, exported. It costs you nothing, it removes the lock-in objection during the sale, and the pipeline was never the valuable part.

How current does the corpus need to be?

Quarterly refresh suits most fields and is what the fee should cover. Anywhere fast-moving, monthly. The important part is showing the date rather than the frequency — a member who knows the corpus stops in June can judge an answer for themselves.