Grounding basics

Grounding means the agent answers from your approved source material instead of guessing.

#How Nexaflow uses knowledge

When a user asks a question, the agent can:

1. Search attached knowledge bases.
2. Retrieve relevant passages.
3. Use those passages as context.
4. Generate an answer constrained by the context and system instructions.
5. Fall back or escalate when the answer is missing.

#Search strategies

Inside Agent Settings → Knowledge Base, choose a search strategy.

Always Search

The agent searches knowledge for every response.

Use this for:

• Customer support.
• HR policies.
• Insurance or finance guidance.
• Legal or compliance-heavy content.
• Public FAQs.

Adaptive Search

The agent decides when to search.

Use this for:

• Conversational lead qualification.
• Sales discovery.
• Lightweight assistants.
• Agents where every message does not need retrieval.

#Multilingual knowledge search

Enable multilingual search when users may ask questions in languages different from your source documents.

Example:

• Source document is in English.
• User asks in Tamil or Hindi.
• Agent searches semantically and replies in the user’s language.

Test multilingual behavior before going live.

#What grounding does not solve

Grounding cannot fix:

• Missing source material.
• Contradictory documents.
• Poorly written policies.
• Wrong module permissions.
• Prompt instructions that override source facts.

If the agent gives a weak answer, improve the source data first.

#Fallback behavior

Good fallback behavior tells the user:

• The agent could not find the specific information.
• What information it can help with.
• Whether the request should be escalated.

Do not instruct the agent to invent facts when knowledge is missing.

#Related guides

• Upload and manage knowledge
• Self-evolving knowledge base
• Testing, sessions, and traces