Agent Instructions vs Knowledge Base

Instructions tell your agent how to behave; knowledge files contain what it retrieves. Keeping them separate is the single biggest factor in answer quality.

Keep instructions and knowledge separate

A common mistake is uploading instructions as knowledge files, or pasting entire documents into the instructions field. Both make it harder for the agent to find the right information.

The two configuration areas

Agent Instructions

Defines role, personality, tone, response format, boundaries, and when to escalate to a human. Always present in the agent's context.

Agent Instructions field in the Basic Info step of agent setup

Where: Basic Info → AI Agent Instructions field

Knowledge Base Files

Documents with factual content the agent searches per question: product docs, policies, FAQs, and crawled web pages.

Knowledge base file upload area for PDFs, documents, and web sources

Where: Agent File Management → Files tab

What goes where

AspectAgent InstructionsKnowledge Base Files
PurposeDefine HOW the agent behavesStore WHAT information it retrieves
ContentRole, tone, rules, escalation triggersFacts, policies, documentation, FAQs
FormatPlain text (roughly 500-1,000 words)PDF, Word, Office, CSV, web sources
How it is usedAlways included in the agent's contextChunked and retrieved per question via hybrid search
UpdatingEdit directly in agent settingsUpload new files, remove outdated ones

How retrieval works

When you upload a file, it is split into small chunks and embedded into the vector database. On each user question, the agent runs a hybrid search - combining keyword and semantic matching - and answers from the most relevant chunks. This is why file structure matters: clear titles, headings, and short paragraphs produce chunks the search can actually match. See how files are organised for file preparation guidelines.

Example: good agent instructions

You are a helpful customer support agent for TechStore, an online electronics retailer.

**Your Role:**
- Assist customers with product questions, orders, and technical support
- Always be polite, professional, and empathetic
- Respond in a friendly, conversational tone

**How to Use the Knowledge Base:**
- Search the knowledge base for product specifications, policies, and FAQs
- Base your answers strictly on retrieved knowledge - never invent facts

**Handling Retrieval Results:**
1. If NO data is found, respond: "I couldn't find that information in our knowledge base. Let me connect you with a team member who can help."
2. If retrieved information CONFLICTS, ask the user to clarify which aspect they are asking about
3. If retrieved information is INCOMPLETE, answer what you found and acknowledge what is missing

**Response Guidelines:**
- Keep responses concise (2-3 paragraphs maximum)
- Use bullet points for lists of features or steps
- For technical issues, provide step-by-step troubleshooting

**Boundaries:**
- Do not make promises about delivery dates without checking the knowledge base
- Do not process refunds or returns - hand off to the returns team
- Do not share internal company information

What makes it good

  • Clear role and tone definition
  • Explicit rules for missing or conflicting data
  • Strict "answer only from knowledge" rule
  • Response format and length guidance
  • Boundaries and escalation triggers

What to avoid

  • Actual product data (put it in knowledge files)
  • Copied policy documents (upload them instead)
  • Instructions longer than about 1,000 words
  • Vague guidance like "be helpful" with no rules

Common mistakes

Pasting facts into instructions

Product catalogs, price lists, and policy documents belong in knowledge files. In instructions they bloat the context and cannot be updated file by file.

Putting behavioral rules in knowledge files

Uploading a PDF called "How to Respond" does not work - knowledge is only retrieved when it matches a user question. Behavior belongs in the instructions field.

No anti-hallucination rules

Without explicit rules for "no data found" and conflicting results, the agent may invent answers instead of admitting it does not know.

Poorly structured knowledge files

Scanned images, missing headings, and huge paragraphs produce fragmented chunks that retrieval cannot match well. Use clear titles, headings, and short paragraphs.

Quick decision checklist

Ask yourselfPut it in
Does it describe how the agent should speak or act?Instructions
Is it a fact a user might ask about?Knowledge file
Does it change often (prices, stock, policies)?Knowledge file
Is it a rule for escalation or refusal?Instructions
Is it longer than a page?Knowledge file

For ready-to-adapt instruction templates, see instructions examples.

Ready to apply this? Create an agent and set up its instructions and knowledge base in a few minutes.

Start building your agent

Last updated · Reviewed by Nemanja Milivojevic, Founder