Split by heading
A document becomes passages that keep their file name and heading, so what reaches an answer is a section with a title rather than an arbitrary slice of text.
Document context
Your CV has one line for the project they will spend ten minutes on.
The architecture, the trade-offs, the numbers and the thing that went wrong are in the design document and the post-mortem. Attach them, and the passage that matches a question is found and used — with the file it came from shown next to the answer.
How it works
Split, then searched per question
- Upload
- Split into passages
- Ranked against the question
- The best ones sent
Everything, when everything fits
Under the budget, all of it is sent in reading order — and there is no retrieval step that can pick wrongly. The clever path only runs when the simple one cannot.
Ranked when it does not
Over the budget, passages are scored against the words of the question and the best are sent. A millisecond, locally, no second model call.
The search
Why word matching and not meaning matching
A deliberate choice, and the opposite of what most products do. Three reasons.
Interviews run on exact terms
“Was that Nomad or Kubernetes”, “which quarter was that”, “was it the Leeds site or Bradford”. The whole question hinges on one token, and a token is what word matching is best at finding — whatever field the token comes from.
A semantic search hedges
Ask about Nomad and a meaning-based search is pleased to offer you the passage about container orchestration in general. That is the wrong passage, confidently chosen.
It costs nothing and no time
No embedding call, no vector store, no second round trip. When the entire budget is four seconds, a millisecond of arithmetic is the right kind of expensive.
Technical tokens survive
“node.js”, “c++”, “s3”, “.net” stay single terms rather than being torn into fragments, because a search that cannot find “c++” is no use in a technical interview.
A relevance floor
A passage has to contain something from the question and score within a fraction of the best one. Padding a prompt with the least-bad passages makes answers worse, not better.
You can see what it used
The file name and heading of each passage are shown with the answer, so a wrong pick is visible to you in a glance rather than inferred from an odd answer.
What to attach
Four kinds of document that earn their place
The test is whether it contains a specific you would not remember accurately under pressure.
Design documents and RFCs
The alternatives you considered and why you rejected them. “Why not use X?” is the follow-up to every architecture answer, and this is where the real reason is written down.
Post-mortems
The incident, the timeline, the root cause, the actions. The strongest material in interviewing and the hardest to recall precisely a year later.
Retrospectives and reviews
What the team changed and what you were credited with. Behavioural answers with a result in them usually come from here.
Anything with numbers
Benchmarks, capacity plans, cost analyses. One figure turns a claim into an answer, and the figures are in these files rather than in your memory.
One caution worth stating plainly: a document can be confidential to an employer whatever you personally think of it. Where it is, write your own summary from memory instead — less detailed, entirely yours, and nearly as useful.
What is sent
The passages that matched, and nothing else
A bounded amount per answer, regardless of how much you keep.
The passages chosen for a question are sent, with the question, to the AI provider writing the answer — the same path your CV and the job description take. Documents you keep but which do not match are not sent. Nothing is used to train anything by us, any document can be deleted on its own, and deleting your account takes all of them with it. The privacy page names the providers and what each may do with what it receives.
Questions
About documents
Why would I attach documents to an interview assistant?
Because a CV has one line for a project an interviewer will spend ten minutes on. The architecture, the trade-offs, the numbers and the thing that went wrong are in the design document and the post-mortem — and those are exactly what a good answer needs and what you will not recall precisely under pressure.
How does it decide which passages to use?
If everything you have attached fits the budget, all of it is sent in reading order and there is no decision to get wrong. When it does not fit, the passages are ranked against the words of the question with BM25 — the ranking a search engine uses — and the best are sent. It runs in about a millisecond, locally, with no extra model call.
Why a word-based search rather than embeddings?
Because interviews are full of exact terms, and exact terms are what word-based ranking is best at. “Did you use Nomad or Kubernetes”, “was that under IFRS 15 or 16”, “which ward was that” — each hinges on a specific token; a semantic search will happily hand back the passage about container orchestration in general. It is also the version that needs no second model call, which matters when the whole budget is four seconds.
Does it tell me which document an answer came from?
Yes. Each passage carries its file name and its heading, and the answer shows what it drew on — so you can see at a glance whether it reached for the right project before you start speaking.
What should I attach?
Design documents and RFCs you wrote, post-mortems for incidents you handled, project retrospectives, a performance review with specifics in it, and anything with numbers in it. What not to attach: anything confidential to an employer, which is a judgement only you can make.
Is attaching my employer's documents a good idea?
Often not, and the decision is yours. A design document may be commercially confidential whatever your own feelings about it. Where it is, write your own summary from memory instead — it is less detailed and entirely yours, and it works almost as well.
How many documents can I attach?
As many as are useful. What is bounded is how much reaches any single answer, not how much you keep — so a large library costs nothing in speed, because the search narrows it per question.
Read next
Related pages
Attach one design document and ask about that project.
It is the quickest way to see the difference between a generic answer and yours.
Create your account