01
Freshness
Use a current evidence set for questions that change over time.
Capability / Current evidence
Live Evidence
Bring current, source-linked evidence into a bounded decision instead of asking a model to improvise from stale context.
The evidence loop
liveSearch the web2 m
Normalize the result18 m
Ask a bounded question3 d
A live context workflow should make the evidence visible before it shows the answer.
The evidence loop
Search first, decide second
A live context workflow should make the evidence visible before it shows the answer.
02
Source weight
Keep the source and retrieval timestamp next to the decision.
03
Bounded answer
Ask a focused yes/no, choice or score question instead of generating an essay.
The evidence loopDecision
Search the web
source 1Collect relevant sources with URLs and retrieval time.
Normalize the result
source 2Put snippets and source metadata into structured state.
Ask a bounded question
source 3Keep the decision narrow enough to inspect and reproduce.
Show the movement
source 4Explain which sources changed or supported the result.
Trust
Current does not mean correct
Fresh evidence still needs source quality, conflict handling and an explicit fallback when the signal is weak.
Ask about current evidence
Current does not mean correct
Fresh evidence still needs source quality, conflict handling and an explicit fallback when the signal is weak.
Ask about current evidence
POST /v1/evidence{
"state": {
"question": "Has the launch date moved?",
"sources": [
{ "url": "https://example.com/notice", "retrieved": "2026-09-29", "text": "Launch moved to October 12." }
]
},
"questions": {
"changed": { "type": "noul", "instructions": "Do the current sources support that the launch date moved?" },
"confidence": { "type": "score", "instructions": "How strong is the evidence?", "criteria": ["weak", "mixed", "strong"] }
}
}FAQ
Questions about live evidence
Is this a general web-search chatbot?+
No. The useful pattern is source collection followed by a bounded decision with visible evidence.
How should conflicting sources be handled?+
Keep the conflict in state, ask about it directly and route low-confidence results to review.
Can I reproduce a decision later?+
Store the source URLs, retrieval time, extracted text, request and response together.
