Ask
Questions answered from your twin, with the rows that answered them
Ask pulls the assumptions, the graph nodes, the run results and the document chunks that bear on your question, and answers from those. Every answer is stored with a list of exactly what was used.
Build a twinWhere a language model is used
With no key configured it returns the retrieved rows instead of a fluent answer. That is deliberate.
A list of the relevant rows is worth more than a fluent sentence with nothing behind it.
The failure mode of a chat box over a business model is an answer that sounds exactly as confident when it is retrieving a real number as when it is inventing a plausible one. Ask is built so that the retrieval is the product and the writing is the optional part, rather than the other way round.
The mechanism
What happens to a question
- 1
The question is reduced to terms
Stop words are dropped and a handful of meaningful terms are kept. No embedding, no vector store, nothing you cannot inspect.
- 2
The ledger is searched
Assumption rows are scored on their key, their label and their note. The best matches come back with their value, unit, origin and the note explaining how they were arrived at.
- 3
The core rows are added regardless
Revenue, gross margin, customer count, churn, headcount, cash and elasticity are included whatever the question was, because almost every question about a company is answered partly by those.
- 4
The graph is searched
Customers, competitors, suppliers and departments whose names match. If nothing matches, the largest few of each type are included instead, so the answer is never empty of context.
- 5
The most recent run is attached
Its scenario, horizon, replication count, the movement in each headline metric against doing nothing, and its most significant events with their frequencies. You can point at a specific run instead.
- 6
The documents are searched
A few chunks from your uploaded files that mention the question's main term, quoted with the file they came from.
- 7
The answer is composed and stored
Along with the citation list: every assumption, node, run and document that went into it, and whether the answer was written by a model or returned as rows.
With no key
It says so, and then it hands you the rows
With no language model configured, Ask does not fail and it does not pretend. It says plainly that no model is configured, lists the retrieved material as it stands, and tells you that adding a key would turn it into prose, or that running the scenario would answer it properly.
This is the behaviour on purpose, not a degraded mode. Most questions people ask a twin are lookups: what is our largest account, what did the last run say about churn, where did this margin figure come from. A list of rows answers all of those completely, and it answers them without a round trip to anybody's API.
It is also the honest floor of the product. If the retrieval is not good enough to be useful on its own, then wrapping it in fluent prose has not made it better, it has only made it harder to check.
- No key, no network call, no cost
- The same retrieval either way, so the material is identical
- Every row labelled with its origin, so a default does not look like a document
With a key
The same rows, written up, under strict instructions
When a model is available it is given the retrieved context and a narrow brief: answer using only this context, quote the numbers you were given, do not compute new ones beyond simple arithmetic on them, and if the context does not support an answer, say which file or which assumption would be needed instead of guessing.
Answers are capped short and kept in plain language. The citation list is stored either way, so you can always see what the answer was built from rather than taking the paragraph on trust.
Every call is cached on a hash of the exact prompt, so asking the same question twice costs nothing the second time.
Four sources
What Ask is allowed to look at
| Source | What comes back | Typical question it answers |
|---|---|---|
| The assumption ledger | Value, unit, origin, band and the note explaining where it came from | Where did our gross margin figure come from, and is it a default |
| The knowledge graph | Nodes with their numeric properties and their share of their own type | Which supplier is the largest share of our spend |
| Runs | Headline movements against doing nothing, and the top events with their frequencies | What did the last price run say about churn in month six |
| Your documents | Chunks that mention the question's main term, with the file name | What did the contract actually say about the most favoured customer clause |
Ask reads your twin and nothing else. It has no access to other accounts, no access to the web, and no general knowledge about your industry that it is willing to offer as fact.
Honestly
What Ask will not do
It is a retrieval tool over one company's model, not an analyst. The boundary is worth knowing before you rely on it.
- It will not run a scenario for you. If the question is about the future, the answer is a run, and Ask will say so
- It will not tell you what your market does next year, because nothing in the twin knows that
- It will not invent a number that is not in the retrieved rows, and with no key it cannot write a sentence at all
- The retrieval is keyword matching on terms, so an oddly phrased question can miss a row that was relevant. The citation list makes that visible rather than hiding it
Ask your own twin where a number came from
The first question most people ask is not about the future. It is about whether the model has understood their company, and that is exactly the question Ask is best at.