Evidence Modes - how strictly your bot uses sources
In the "Anchor sources" node you decide how strictly your bot sticks to the knowledge snippets it finds. There are three modes to choose from: Strict, Balanced and Helpful (no hallucination). Chosen wrongly, the bot becomes either too reserved or too creative. This guide helps you find the right mode.
TL;DR
- Strict: Only statements with a clear source reference. Frequent "I don't know" answers.
- Balanced (default): Sources are preferred, but the bot fills gaps politely.
- Helpful: The bot answers even without perfect sources, but states what is uncertain.
- Minimum sources reinforce the respective mode.
- The wrong mode is one of the most common sources of hallucination.
What the modes mean technically
When your bot receives a question, it searches for knowledge in your database. The hits land in the answer prompt - but how closely the bot sticks to them is controlled by the evidence mode.
| Mode | Behavior |
|---|---|
| Strict | Every statement must be backed by a source. For gaps, the bot says "I don't know that". |
| Balanced | Sources are preferred, but the bot can supplement with general knowledge when it is plausible. |
| Helpful | The bot always answers, but transparently states when it does not know something from sources. |
When "Strict" fits
Strict is the right choice when wrong answers are more expensive than no answer at all:
- Compliance bots - legal topics, data protection, insurance.
- Medical bots - also in a broad sense, for example with symptom questions.
- Technical specification bots - where values or thresholds must be exactly right.
- B2B sales bots - where a wrong pricing statement alienates a customer.
In these cases, combine Strict with:
- Minimum sources = 2 (at least two hits required).
- Minimum similarity >= 0.60 (only very matching hits count).
- Top-K = 6 or 8 (smaller pool, clearer hits).
Expect the bot to say "I don't know that" on 20 to 30 percent of all questions. That is intended - it should rather stay silent than claim something false.
When "Balanced" fits
Balanced is the default and fits 80 percent of all bots:
- FAQ bots where end users expect a concrete answer.
- Product information bots with a good but not seamless knowledge base.
- Coaching and advisory bots where human empathy matters more than scientific rigor.
In this mode, the bot mixes source knowledge and general knowledge - what it knows comes from sources, what it does not know it phrases with caution. The mode is typically run with minimum sources 1 and threshold 0.50.
When "Helpful" fits
Helpful is the softest mode. The bot always answers but points out uncertainties:
- Marketing bots and demo bots, where conversation matters more than accuracy.
- Personality bots with a strong persona, where knowledge is secondary.
- Lore wikis and fan bots, where inventiveness is desired.
The Helpful mode will certainly produce hallucinations - that is part of the design. Anyone who cannot accept that should choose "Balanced".
Minimum sources - the lever behind the mode
In the inspector you find minimum sources with the values 0, 1, 2, 3:
| Value | Effect |
|---|---|
| 0 | Bot answers even without hits, dangerous in "Strict" |
| 1 | Default, one hit is enough |
| 2 | Recommended for "Strict", clearly reduces hallucinations |
| 3 | Very restrictive, only for compliance bots |
If fewer than the minimum number was found, the fallback is triggered. The default fallback is "I don't know" - alternatively "Hand off to a human" or "Answer anyway".
How to set this in the Visualizer
- Open the Visualizer and click the "Anchor sources" node.
- In the inspector you choose:
- Evidence mode: Strict | Balanced | Helpful
- Minimum sources: 0-3
- Fallback: Answer when there are not enough hits
- Save.
- Test with three questions:
- One whose answer is clearly in your knowledge base.
- One that is only partially covered.
- One that is not covered at all.
Observe whether the behavior matches the chosen mode. In Strict mode the third question should trigger an "I don't know". If it does not, the knowledge base is too broad or the threshold is too low.
Common mistakes
- Strict + minimum sources 0: Contradiction. The bot should be strict, but answer without sources? Minimum sources at least 1, better 2.
- Helpful + minimum sources 3: Also a contradiction. If you use Helpful, you always want an answer - set minimum sources to 0 or 1.
- Mode switch without testing: A reconfigured pipeline without a subsequent live-preview run often gets worse.
- Strict for small-talk bots: Greetings, thank-you phrases and follow-up questions should not force the bot into "I don't know" - combine Strict with the activated "Small-talk switch" node.
For advanced users
In Strict mode the persona style is partly overridden - the bot quotes more faithfully to the source, less in the persona tone. If your persona is strong but Strict is necessary, compensate in the persona text with an instruction like:
Always answer in the defined style, even when you only answer briefly or with "I don't know".
Next steps
- Understand Precision vs Recall to set Top-K and threshold to match the mode choice.
- Check your answer behavior and limits - many mode effects are connected to it.
- If the bot stays silent too often in Strict, Increase the hit rate on questions also helps.