Grounded answers API
Siltstone reads a text once, then answers thousands of questions about it: yes or no, which one, how much, when. Every answer comes typed, with a probability your code can act on.
No card needed · $0.042 per million input tokens · answers are free
Inbox · support@
Dana Whitfield
Charged twice. Still waiting.
Hi team, we were charged twice for our March invoice (INV-2291, £1,840.50, issued 03/03/2026) and nobody has replied to my last two emails. This is the third billing problem this quarter. Please get a manager to call me today, otherwise we'll have to look at other providers. Dana Whitfield, Operations Lead, Northwind Freight
8 questions ready
{"document": "Hi team, we were charged twice for our March invoice (INV-2291, £1,840.50, issued 03/03/2026) and nobody has replied to my last two emails. This is the third billing problem this quarter. Please get a manager to call me today, otherwise we'll have to look at other providers. Dana Whitfield, Operations Lead, Northwind Freight","questions": [{"id": "urgent","type": "bool","ask": "Does this need a reply today?"},{"id": "team","type": "choice","ask": "Which team should handle it?","options": ["billing","technical","sales"]},{"id": "refund","type": "bool","ask": "Are they asking for money back?"},{"id": "manager","type": "bool","ask": "Do they want a manager to call?"},{"id": "churn","type": "score","ask": "How likely are they to leave?","levels": ["Happy","Complaining","Comparing alternatives","Will leave unless fixed"]},{"id": "amount","type": "number","ask": "How much was the invoice, in pounds?","decimals": 2},{"id": "issued","type": "date","ask": "When was the invoice issued?","order": "day_first"},{"id": "repeat","type": "number","ask": "How many billing problems this quarter?"}]}
{"document": "Hi team, we were charged twice for our March invoice (INV-2291, £1,840.50, issued 03/03/2026) and nobody has replied to my last two emails. This is the third billing problem this quarter. Please get a manager to call me today, otherwise we'll have to look at other providers. Dana Whitfield, Operations Lead, Northwind Freight","questions": [{"id": "urgent","type": "bool","ask": "Does this need a reply today?"},{"id": "team","type": "choice","ask": "Which team should handle it?","options": ["billing","technical","sales"]},{"id": "refund","type": "bool","ask": "Are they asking for money back?"},{"id": "manager","type": "bool","ask": "Do they want a manager to call?"},{"id": "churn","type": "score","ask": "How likely are they to leave?","levels": ["Happy","Complaining","Comparing alternatives","Will leave unless fixed"]},{"id": "amount","type": "number","ask": "How much was the invoice, in pounds?","decimals": 2},{"id": "issued","type": "date","ask": "When was the invoice issued?","order": "day_first"},{"id": "repeat","type": "number","ask": "How many billing problems this quarter?"}]}
How it works
Reading the text is the expensive part. Siltstone does it once, then every question runs against that single read, costing only its own few tokens. Ask more, and each answer gets cheaper.
238 questions
Time
Cost
Faster
55×
than one LLM call
Cheaper
132×
than one LLM call
Per 1,000 answers
$0.0012
826,527 answers for $1
Siltstone times on a single H100 serving the current model, at $0.042 per million input tokens. The LLM is an estimate: every question in one prompt, typical frontier prices ($1.25 in, $10 out per million tokens) and speeds (0.4 s to the first token, then ~80 tokens a second of JSON answers).
Compared
Put fifty questions in one LLM prompt and they compete for its attention: answers drift, and 'no' quietly turns into 'yes'. Siltstone answers every question on its own against the same read, so the hundredth is as good as the first.
Hover or tap a column for its numbers.
62× cheaper
One 4,000-token text, every question in one request. The LLM is assumed at typical frontier prices, $1.25 in and $10 out per million tokens, with answers returned as JSON; Siltstone at $0.042 per million, answers free.
Benchmarked over 12 documents of ~14,000 tokens, every question sent in one schema-enforced call to a popular frontier model. At 24 questions: 94.8% for Siltstone against 80.2% for the single call.
Question types
No prose to parse. A bool comes back true or false, a choice as your option's name, a number as a number and a date as an ISO string, each in its own type, each with its confidence.
The text
Survey of mast NW-4471, completed 14/06/2025 at 09:20 by R. Okafor. Operator: Vodafone. Height to the top of the antenna: 22.4 m. Three sectors, bearings 40°, 160° and 280°. Surface rust on two brackets; structure sound. No lightning protection fitted.
Does the text say it? True or false, with the probability on each.
{"type": "bool","ask": "Is lightning protection fitted?"}
Confidence
Every answer carries the probability the model put on it. Set a threshold on the request, and each answer below it comes back flagged for review, so the rule in your code is one line.
35acted on automatically
5flagged for a person
// the whole rule
if (answer.review) sendToPerson(answer); else act(answer.value);Illustrative answers. Where to draw the line depends on your texts: measure it on a few dozen you know the answers to.
For developers
Send the text and every question you have about it. Branch on the values; send the flagged ones to a person.
curl https://api.siltstone.ai/v1/ask \
-H "Authorization: Bearer $SILTSTONE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"document": "Charged twice for INV-2291 (issued 03/03/2026). Fix it today or I cancel.",
"questions": [
{ "id": "urgent", "type": "bool", "ask": "Does this need a reply today?" },
{ "id": "team", "type": "choice", "ask": "Which team?",
"options": ["billing", "technical", "sales"] },
{ "id": "invoice_date", "type": "date", "ask": "When was the invoice issued?" }
]
}'Response · 200
{"threshold": 0.7,"flagged_for_review": 0,"answers": [{ "id": "urgent", "type": "bool", "value": true,"confidence": 0.97, "review": false },{ "id": "team", "type": "choice", "value": "billing","confidence": 0.96, "review": false,"probabilities": { "billing": 0.96, "technical": 0.03, "sales": 0.01 } },{ "id": "invoice_date", "type": "date", "value": "2026-03-03","parts": { "year": 2026, "month": 3, "day": 3 },"confidence": 0.96, "review": false }],"total_ms": 48.2}
Values, not prose
true, "billing", 1840.5, "2026-03-03". The type is the contract, so nothing needs parsing.
Up to 2,000 questions
All against one read of the text, in one request. Numbers and dates are written under a grammar, so they always parse.
A probability on everything
Set a threshold and the unsure answers flag themselves. How to read one.
Pricing
$0.042 per million input tokens: your text and your questions. Prepaid credit, no subscription, and $5 to start.
What would it cost me?
A month
$14.45
Per document
$0.00014
3,440 tokens
Your free $5
34,606
documents like these
$0.042
per million input tokens
What you can build
Support
“The export button charged us twice and the file never downloaded. Finance needs it before 3pm today.”
→ Open a technical ticket, due 15:00; flag billing.
Sales
“New workspace: Northline Studio. 14 teammates invited, 3 projects, and an email asking about SSO and annual invoicing.”
→ Send the account to the team-plan queue.
Calls
“This is the third time I've called. The replacement still hasn't arrived. Please let me speak to a supervisor.”
→ Surface the escalation to the agent, live.
Finance
“Invoice 2291 from Harbour Freight Ltd, dated 3 March 2026, total £1,840.50 inc. VAT, due within 30 days.”
→ Book it, schedule payment for 2 April.
Search
“Question: can I return a sale item? Passage: sale items may be returned within 14 days if unworn. Final-sale items are excluded.”
→ Keep the passage; pass the conditions on.
Agents
“Task: report duplicate files. Proposed action: delete 214 files. Policy: read-only unless a person approves.”
→ Hold the action for approval.
Start free
Create an account, make a key, and ask your first thousand questions before you've thought about a card.