Every token,
traced to its source.
TokenPath shows exactly where each span of an LLM answer came from — character-level offsets into your documents, with confidence scores. One API call, any model's output.
In Q3, Northwind's revenue grew 18% year over year to $42.1M, driven primarily by the enterprise segment. Gross margin expanded to 71%, up from 66% a year earlier. The company ended the quarter with $310M in cash and equivalents, and management raised full-year guidance to a range of $168–172M.
Q · How fast did revenue grow, and why?
Attention in, evidence out.
Send what the model read, and what it wrote.
One request: the source document, the question, and the answer any decoder LLM produced. No retraining, no logprobs, no vendor lock on the generating model.
POST /v1/attributions { "document": "In Q3, Northwind's revenue grew 18%…", "question": "How fast did revenue grow?", "answer": "Revenue grew 18% year over year.", "threshold": 0.001 }
Works with output from any model — yours, OpenAI's, Anthropic's.
We trace attention through the model.
TokenPath runs an instrumented forward pass and reads where the model actually looked — token-level attention, distilled down to character-level evidence. You can even steer which layers and heads to trust.
Get spans back, with scores.
For every answer span: exact character offsets into your document, plus a confidence score. Render citations, gate hallucinations, keep audit receipts.
{
"source_span": { "start": 412, "end": 471, "score": 0.97 }
}…In Q3, Northwind's revenue grew 18% year over year to $42.1M, driven primarily by…
POST /v1/attributions { "document": "In Q3, Northwind's revenue grew 18%…", "question": "How fast did revenue grow?", "answer": "Revenue grew 18% year over year.", "threshold": 0.001 }
Works with output from any model — yours, OpenAI's, Anthropic's.
{
"source_span": { "start": 412, "end": 471, "score": 0.97 }
}…In Q3, Northwind's revenue grew 18% year over year to $42.1M, driven primarily by…
Ship citations in an afternoon.
No SDK required to start — it's one HTTPS endpoint. Send the document, the question, and the answer; get spans back. Rate limits, usage metering, and key management come with the platform.
# one call, any model's output curl https://api.tokenpath.ai/v1/attributions \ -H "Authorization: Bearer $TOKENPATH_API_KEY" \ -d '{ "document": "In Q3, Northwind's revenue grew 18%…", "question": "How fast did revenue grow?", "answer": "Revenue grew 18% year over year." }'
# spans with character offsets + scores { "source_span": { "start": 412, "end": 471, "score": 0.97 }, "text": "revenue grew 18% year over year to $42.1M" }
Trust is a feature. This is its API.
Citations that point at sentences, not chunks.
Turn retrieved-context answers into precisely cited ones. TokenPath returns exact character ranges, so highlights land on the words that carried the answer — not a 500-token chunk.
Gate hallucinations before users see them.
Spans with no supporting source come back with low confidence. Set a threshold, flag or block unsupported claims, and ship answers you can stand behind.
Keep receipts for every generated word.
Regulated teams need to show where an answer came from. Store TokenPath spans alongside outputs and every response carries its own audit trail.
Debug models by watching where they look.
Groundedness metrics for eval suites, attention-level diagnostics for model work. Steer layers and heads to see exactly what the model attended to.
A dollar per million tokens. That's the pricing page.
Usage-based, no monthly minimum, 1M tokens free when you sign up. Volume pricing and dedicated deployments for enterprises.
See pricing detailsStop asking users to trust the model.
Show them the source.
1M free tokens · no card required