Teach your assistant how to prioritize information, avoid hallucination, and deliver precise answers
You have uploaded your pricing, policies, and product docs. You have written clear instructions. But your assistant still gives vague answers, mixes up information from different sources, or occasionally makes things up.The missing piece is context engineering: teaching your assistant how to find, prioritize, and use information from multiple sources. This is what separates assistants that “kind of work” from assistants that feel like a real team member.
Context engineering is the practice of telling your assistant which information to trust, in what order, and how to handle gaps. It lives inside your Instructions and works together with your Knowledge Base.Without context engineering, your assistant treats all Knowledge equally. A blog post from six months ago carries the same weight as your current pricing page. An outdated FAQ competes with your latest policy update.With context engineering, your assistant knows:
Which sources to check first
How to handle conflicting information
What to do when it cannot find an answer
How to present sources naturally
Context engineering is especially important once your assistant has more than 2-3 Knowledge items. The more sources you add, the more your assistant needs guidance on how to use them.
The most impactful thing you can add to your instructions is a trust order for your knowledge sources. This tells the assistant which source wins when information overlaps or conflicts.Add a section like this to your Instructions:
## Knowledge Source PriorityWhen answering questions, check sources in this order:1. Pricing documents and pages (for costs, plans, and billing)2. Official policies (for refunds, terms, and cancellations)3. Product documentation and FAQs (for features and troubleshooting)4. Announcements and changelogs (for recent updates)If two sources give different answers, use the most recentlyupdated source. If you are still unsure, tell the customer andoffer to connect them with our team.
This prevents a common failure: the assistant quoting an outdated FAQ when the pricing page has the correct, current number.
Guide how your assistant constructs its answers from knowledge sources.
## How to Answer- Find the specific, relevant section in the knowledge base before answering- Lead with a direct answer in 1-3 sentences- Add supporting details (steps, bullet points) only if needed- If a key detail is missing from the conversation, ask one clarifying question before making assumptionsExample: Ask "Which plan are you interested in?" instead ofguessing the customer's plan.
This produces focused, accurate responses instead of long, wandering answers that try to cover everything.
This is the most important safety net in your instructions. Tell your assistant exactly what to do when it does not have the answer.
## When You Don't KnowIf the knowledge base does not contain the answer:- Never guess, estimate, or invent information- Say: "I don't have that information in our docs yet. Would you like me to connect you with our team?"- For pricing: never make up numbers. Say: "Let me check with our team on that specific pricing."
Without this rule, AI models will try to be helpful by generating plausible-sounding answers that may be completely wrong. An explicit fallback prevents this.
Hallucination is the most common reason businesses lose trust in their AI assistant. A clear “I don’t know” is always better than a confident wrong answer. Make your fallback instructions specific and direct.
When your assistant references information from its Knowledge Base, the default behavior can sound robotic. Teach it to share sources naturally.
## Sharing SourcesWhen referencing information from the knowledge base, usenatural language instead of formal citations:Good:- "You can find the full details on our pricing page"- "Our refund policy covers this. Here's the key part..."- "More info here if you'd like to read the full breakdown"Bad:- "Source: pricing.pdf"- "According to document #3..."- "Reference: company-policies.docx, page 12"Only link to URLs that actually exist in the knowledge base.Never generate or guess URLs.
As your Knowledge Base grows, you may end up with information that overlaps or contradicts itself. Prepare your assistant for this. Keep your Knowledge Base well-organized with one authoritative source per topic to avoid conflicts in the first place.
## Conflicting InformationIf you find conflicting details across different knowledge sources:1. Use the source highest in the priority order2. If both sources are at the same level, use the one that was most recently updated3. If you cannot determine which is correct, tell the customer: "I'm seeing some conflicting details on this. Let me connect you with our team to make sure you get the right answer."
The best way to prevent conflicting information is to audit your Knowledge Base regularly. Remove outdated documents, merge overlapping content, and keep one authoritative source for each topic. See the maintenance tips below.
Here is a full context engineering section you can add to your Instructions:
## Context Engineering### Source Priority1. Pricing and billing documents2. Official policies (refunds, terms, cancellations)3. Product docs and FAQs4. Blog posts and announcements### Response Method- Always check the knowledge base before answering factual questions- Lead with a direct, short answer- Add details only when helpful- Ask one clarifying question if key information is missing### When You Don't Know- Never guess or make up information- Say: "I don't have that in our docs yet. Want me to connect you with our team?"### Conflicting Information- Follow the source priority order above- When unsure, escalate to a human### Source References- Use natural language when referencing sources- Never show raw file names or document IDs- Only link URLs that exist in the knowledge base
Customer: “What’s the refund policy for the Pro plan?”Assistant: “Refunds are generally available within 30 days.” (vague, possibly wrong, no source)
With context engineering:
Customer: “What’s the refund policy for the Pro plan?”Assistant: “Pro plan subscriptions can be cancelled anytime, but refunds are only available within the first 14 days of a billing cycle. After that, your access continues until the end of the current period. You can find the full policy details here if you’d like to review them.”
The difference: the assistant checked the right source, gave a precise answer, and referenced it naturally.
Context engineering is only as good as the Knowledge Base it works with. Keep both healthy:Weekly: Review conversation logs in the Inbox for questions the assistant struggled with. Check whether the answer exists in Knowledge and whether the assistant found it.Monthly: Audit your Knowledge Base for outdated or conflicting content. Remove documents that have been superseded. Merge overlapping sources into one authoritative version.After business changes: When you update pricing, policies, or products, update the corresponding Knowledge items immediately. Outdated knowledge with good context engineering still produces wrong answers.
For websites added to Knowledge, use the three-dot menu and select “Re-index” to refresh the content after your site changes. For files and text, you will need to reupload or rewrite them with updated information.
