How Ask Ai works

Last updated 17 September 2026

Ask Ai is a research assistant for SEND. You ask a question in plain English, it finds the relevant pages in the Library and writes a short answer with real quotes and citations for every claim. This page explains how it works. You can also read our AI statement.

The seven layers of AI

Layer		What it is	Ownership	Location
Interface	The bit you see and type into	The question box at poppop.studio/tools/ask/. Three questions free, 50 a day once signed in. Built by us.	Pop Pop Studio	France
Harness	Our code that runs the whole job, from your question to a checked answer.	Our own code and prompts. It runs the whole job in order: takes your question, calls a model to check it back to you and widen it, runs the search, sends the pages to the model that writes the answer, then checks every quote in that answer against its source. Built and owned by us, on our server.	Pop Pop Studio	France
Retrieval	Finding the right pages in the Library before any model sees anything	Every document in the Library is split into paragraphs of about 150 words, roughly one real point each. Each paragraph is turned into a list of numbers that captures its meaning (an "embedding"), using a small open model called bge that runs on our own server. Those numbers, with the document title, page number and topic tags, go into a search index (Qdrant) on the same server. When you ask a question, it is turned into numbers the same way. The index returns the paragraphs whose meaning is closest, not the ones that share the most words. Two rules keep it fair: no single document may fill more than two of the slots, so one dense document cannot crowd out the others; and each matched paragraph is expanded back to its whole page, so the model reads the point in context rather than an orphaned sentence. Up to 15 pages go forward. The model never sees anything the search did not find.	Pop Pop Studio	France
Routing	Deciding which model does which of the small jobs along the way	One question makes four calls to a model. Three are small, and go to Claude Haiku one checks the question back to you and judges how long a good answer should be (60 to 120 words for a factual question, up to 500 for a broad one) one breaks the question into four or five search angles so the search covers the topic one reorders the found pages by how useful they are for answering, not just how similar they look. The fourth call is the big one, to Claude Sonnet: it reads the pages and writes the answer with citations. Which model does which job is a setting, not code. Any of the four could move to another model or another host on its own, and the 16 evaluation questions would say whether it was a good move.	A choice, currently Anthropic	USA
Model	The trained system that reads the pages and writes the words.	Claude Sonnet 5 writes the answers and Claude Haiku 4.5 does the three small jobs, both made by Anthropic and chosen on results. The embedding model that powers the search is different: BAAI bge-base, an open model, and it runs on our server in France.	A choice, currently Anthropic Pop Pop Studio	USA France
Inference host	The company, and country, whose computers the model actually runs on.	Anthropic, United States. Your question and the source pages go there, are not used for training, and are deleted within 30 days. The transfer is covered by the UK International Data Transfer Addendum. We keep your question and answer on our own server for 12 months.	A choice, currently Anthropic	USA
Evaluation	How we know it still works after we change something	16 real SEND questions with marked answers, run blind against any model on any host, plus an automatic check that every quote exists word for word in the source it is credited to. This is how we know a change has not made it worse, and how a switch of model or host is decided. Tested: Llama 3.3 70B running on OVHcloud, and models hosted by Scaleway in Paris, have been put through the same harness and the same 16 questions. For word for word quotation Claude is still clearly better, so it is the one in use while the product is being built. Once we are properly in production we will review this, and the review is a re-run of the test questions, not an opinion.	Pop Pop Studio	France

What we have found so far

In the September 2026 blind test Claude produced 95 quoted citations and 7 per cent could not be verified. Llama 3.3 70B produced 8 quotes and half could not be verified. Both mislabelled real sources as "the SEND Code of Practice" on one question. For word for word quotation the US model is still clearly better, which is why it is the one in use. For plain search and summarising, the open models are probably already good enough. We keep testing.

The checker has caught the model paraphrasing a sentence and presenting it as a direct quote, and a real quote credited to the wrong source. Both would have passed without it.