Show HN: TokenPath – token-level citations for LLM output, read from attention
9 hours ago
- TokenPath provides token-level attribution for AI model answers, resolving each claim to the exact source span in the document.
- It works with any model's output without changing the stack, requiring only a single API call after the answer is generated.
- Three citation methods are compared: ask model (no granularity), retrieve and re-rank (cannot disambiguate identical values), and read attention (granular, unambiguous, and fast).
- On LongBench-Cite, TokenPath's read-attention method achieves 0.815 F1, matching Anthropic's Citations API at about 7× lower cost and 5–6× faster speed.
- The API accepts a document, question, answer, and claim spans, returning a source span and confidence score for each claim.
- Pricing is $1 per million tokens with 10 million free tokens upon signup, and no monthly minimum.
- TokenPath is designed for the agentic era, enabling auditing of what agents actually used by reading attention from the model.