Confidence & Evidence Model
Every AI output in Virza carries transparency indicators so you can assess reliability before using results in your research. This page explains what those indicators mean and how they work.
Knowledge source badges
Every AI chat response shows a colored badge indicating where the information came from:
| Badge | Color | Meaning |
|---|---|---|
| ”Based on X sources from your workspace” | Green | Answer is fully grounded in your uploaded documents. Every claim is backed by a citation. |
| ”X workspace sources + general knowledge” | Violet | Answer blends evidence from your documents with the AI model’s training knowledge. Cited content has green borders; general knowledge has amber borders. |
| ”General knowledge” | Amber | No relevant documents were found in your workspace. Answer comes entirely from the AI model’s training data. |
For formal research outputs, rely on green-badge responses. Violet and amber responses are useful for exploration but should be verified against primary sources before citation.
Evidence strength tiers
When Virza cites a source from your library, each source card shows an evidence strength indicator:
| Tier | What it means |
|---|---|
| Strong | High semantic similarity between your question and the cited passage. The passage directly addresses what you asked. |
| Good | Meaningful relevance. The passage discusses the topic but may not directly answer the specific question. |
| Weak | Tangential relevance. The passage mentions related concepts but may require interpretation. |
Evidence strength is determined by combining the neural reranker’s cross-attention score (how deeply relevant the passage is to your query) with the vector similarity score (how close the meaning is). Strong evidence requires high scores on both signals.
Confidence meter
Responses grounded in your documents show a confidence bar:
- 90–100%: Nearly all claims in the response are supported by cited evidence
- 70–89%: Most claims are supported; some inferences are made to connect evidence
- Below 70%: Significant portions of the response rely on general knowledge or inference
The meter measures the ratio of claims that have direct citations to total claims in the response.
Evidence sufficiency warnings
Before you read a response, Virza checks whether the retrieved evidence is sufficient:
- No warning: strong, comprehensive evidence found
- Amber badge: “Limited evidence”: some relevant documents found, but coverage may be incomplete
- Red badge: “Insufficient evidence”: very few or no relevant documents found; the response will rely heavily on general knowledge
Inline provenance
In hybrid responses (violet badge), visual markers distinguish source types:
- Green left border: this text is directly supported by cited documents from your workspace
- Amber left border: this text comes from the AI model’s general knowledge
This lets you quickly scan a response and identify which parts are evidence-backed.
Transparency settings
Control how much detail Virza shows about its reasoning. Open the AI Profile controls (sliders icon in chat) → Advanced → Transparency:
| Setting | What you see |
|---|---|
| Minimal | Knowledge source badge only. Clean and distraction-free. |
| Standard (default) | Badge + evidence strength tiers on source cards + warnings when evidence is weak. |
| Verbose | Full research mode: badge + strength + confidence meter + inline provenance borders + auto-expanded deep research reasoning chain. |
For systematic reviews and formal research, use Verbose transparency. For everyday reading and exploration, Standard provides the right balance.
What this means for your research
- Cite with confidence: green-badge responses with Strong evidence are suitable for use in academic work. Always verify the original source passage via the source card link
- Use amber responses as leads: general knowledge responses can point you toward topics to investigate but should not be cited directly
- Check the reasoning chain: in Deep Research mode, the collapsible timeline shows every evidence retrieval hop, what gap was identified, and what was retrieved, giving you full auditability
- Compare across sources: when multiple source cards are shown, check whether they agree or present different perspectives
Limitations
- Confidence reflects retrieval quality, not factual accuracy: a Strong match means the passage is highly relevant to your question, not that its content is correct
- The AI model may misinterpret passages even with high retrieval scores. Always verify important claims against the original document
- Very recent research (published in the last few days) may not yet be indexed in external databases on the Discover tab
- Confidence meters are based on citation counting, not semantic verification of each claim’s correctness
- Cross-document synthesis may introduce inferences that aren’t in any single source. The confidence meter accounts for this by scoring synthesized claims lower