Verifiable Search

Verifiable generative search

Generative search answers can sound confident while citing weak, mismatched, or fabricated evidence. This app makes every answer show the exact source text behind its citations, catches unsupported citations, and tracks how often verification passes or fails.

Benchmark goal: measure citation accuracy across saved runs and compare against best-in-class generative search benchmarks.

Verifiable Search

Your favorite search engine, now backed by verifiable sources.

Exa

Verifiable Search

Every highlighted word is checked against its source.

AI search answers can be fabricated, or backed by citations that lead nowhere. Verifiable Search keeps the reach of a normal search engine and adds SourceryKit on top, so you can check any claim against the text it came from.

Click the highlight to see the source quote behind it.

The James Webb telescope launched on December 25, 2021 from French Guiana.

What SourceryKit does

SourceryKit is a cryptography SDK for runtime verification. It:

captures the sources an agent checked and the data it received
holds the model to answers that match the retrieved data
returns answers backed by real, verified citations

It drops into any agentic tool-calling workflow: enterprise search, RAG pipelines, MCP servers, APIs, multi-agent systems.

How a search runs

1 Search Exa retrieves live web sources.
2 Draft The model writes an answer from those sources.
3 Check SourceryKit checks each claim against exact quotes.
4 Show Passed claims become evidence chips you can click.

Why this matters

Independent audits of AI search:

Perplexity led the Tow Center benchmark and still answered incorrectly 37% of the time.
ChatGPT Search misidentified 134 of 200 news articles in the same benchmark. UPenn measured 10.1% non-resolving URLs in OpenAI Deep Research, 3.5% of them likely invented.
Gemini returned one fully correct response in the Tow benchmark. BBC and EBU found sourcing problems in 72% of its answers.
Claude sat out the Tow benchmark. UPenn measured 7.8% to 8.5% non-resolving URLs, with few invented ones.

Answer

The recommended pattern is to search first, fetch source content, extract answerable evidence, and only then ask the model to compose a response with citations. The app should preserve both the normalized source text and the exact evidence snippets used for each claim. 1 2

For user trust, every answer should expose a verification record: query, search provider request id, selected URLs, fetched timestamps, content hashes, claim-to-source mappings, and verification status. 3

Verifiable generative search