> ## Documentation Index
> Fetch the complete documentation index at: https://www.meilisearch.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Reduce hallucination
> Techniques to minimize LLM hallucination in conversational search, including few-shot prompting, system prompt engineering, and guardrail strategies.
Large language models can sometimes generate information that is not present in the source documents. This is known as hallucination. While it cannot be fully eliminated, several techniques significantly reduce its occurrence in Meilisearch's conversational search.
## Understanding why hallucination happens
When Meilisearch sends retrieved documents to the LLM, the model may:
* Fill gaps in the provided context with its own training data
* Misinterpret ambiguous information in the documents
* Combine facts from different documents in incorrect ways
* Generate plausible-sounding but fabricated details
The key principle is: **the LLM should only use information from the documents Meilisearch retrieves, not its general knowledge**. By default, LLMs may draw on their training data to fill gaps. All the techniques below help enforce this boundary and keep responses grounded in your indexed data.
## System prompt engineering
The system prompt is your first and most important line of defense. A well-crafted system prompt sets clear boundaries for the model.
### Be explicit about data boundaries
```text System prompt theme={null}
You are a search assistant for our product documentation. You MUST
follow these rules:
1. Only answer using information from the search results provided
2. If the search results do not contain enough information to answer
the question, respond with: "I could not find this information in
our documentation."
3. Never use your general knowledge to fill in gaps
4. Never invent product features, prices, or specifications
```
### Specify how to handle uncertainty
```text System prompt theme={null}
When you are not confident in your answer:
- Say "Based on the available documents, ..." to signal partial
information
- List what you found and what is missing
- Suggest the user refine their search query or contact support
```
### Require source attribution
Forcing the model to cite its sources makes it harder to hallucinate, because fabricated information has no source to point to:
```text System prompt theme={null}
For every claim in your answer, reference the specific document it
comes from. Use the format [Source: document title]. If you cannot
attribute a claim to a specific document, do not include it.
```
## Few-shot prompting
Few-shot prompting provides the model with examples of correct behavior directly in the system prompt. This is one of the most effective techniques for reducing hallucination.
### Show the model what good answers look like
Include 2-3 examples in your system prompt that demonstrate the expected behavior:
```text System prompt theme={null}
You are a product support assistant. Answer questions using only the
search results provided. Here are examples of how to respond:
Example 1:
User: "What is the battery life of the X100?"
Search results contain: "The X100 features a 4500mAh battery with
up to 12 hours of screen-on time."
Good answer: "The X100 has a 4500mAh battery that provides up to
12 hours of screen-on time. [Source: X100 product page]"
Example 2:
User: "Does the X100 support wireless charging?"
Search results contain: Information about X100 battery and display,
but nothing about wireless charging.
Good answer: "I could not find information about wireless charging
for the X100 in our documentation. You may want to check the full
specifications page or contact our support team."
Example 3:
User: "Compare the X100 and the Y200 battery life."
Search results contain: "X100: 4500mAh, 12h screen-on time" and
"Y200: 5000mAh battery"
Good answer: "The X100 has a 4500mAh battery with up to 12 hours
of screen-on time. The Y200 has a larger 5000mAh battery, but I
could not find its screen-on time in the available documents.
[Sources: X100 product page, Y200 product page]"
Now answer the user's question following this same pattern.
```
### Show the model what to avoid
Negative examples are equally powerful. Show the model what a hallucinated answer looks like:
```text System prompt theme={null}
NEVER respond like this:
User: "Does the X100 support 5G?"
Search results: No mention of 5G.
Bad answer: "Yes, the X100 supports 5G connectivity with sub-6GHz
and mmWave bands."
This is wrong because the answer fabricates information not in the
search results.
```
## Guardrails in Meilisearch Cloud
Meilisearch Cloud provides built-in guardrail options through the workspace settings. These guardrails work by injecting carefully crafted instructions into the system prompt to guide the model's behavior.
Guardrails are prompt-based, meaning they shape the model's behavior through instructions rather than through hard technical constraints. They significantly improve response quality but should be combined with monitoring for production use.
Configure guardrails through the [chat workspace settings](/capabilities/conversational_search/how_to/configure_chat_workspace) or the Meilisearch Cloud UI. Available guardrails include:
* **Scope restriction**: limits the topics the agent discusses
* **Data grounding**: forces the agent to only use retrieved documents
* **Response formatting**: controls the length and structure of answers
For detailed configuration examples, see the [Configure guardrails](/capabilities/conversational_search/how_to/configure_guardrails) guide.
## Combine techniques for best results
In production, use multiple techniques together. Here is an example of a system prompt that combines system prompt engineering, few-shot prompting, and source attribution:
```text System prompt theme={null}
You are the documentation assistant for CloudDeploy. Answer questions
using ONLY the search results provided by Meilisearch.
Rules:
- Never use information from your training data
- Cite sources for every claim: [Source: document title]
- If you cannot find the answer, say "I could not find this in our
documentation" and suggest contacting support@clouddeploy.com
- Keep answers concise (under 150 words) unless more detail is
requested
Example of a good answer:
User: "How do I configure auto-scaling?"
Search results: "Auto-scaling can be enabled in the dashboard under
Settings > Scaling. Set min and max instances."
Answer: "To configure auto-scaling, go to Settings > Scaling in the
CloudDeploy dashboard. There you can set the minimum and maximum
number of instances. [Source: Auto-scaling configuration guide]"
Example of handling missing information:
User: "What are the pricing tiers?"
Search results: No pricing information found.
Answer: "I could not find pricing information in our documentation.
Please visit our pricing page or contact support@clouddeploy.com
for current pricing details."
```
## Monitor and iterate
No prompt configuration is perfect from the start. Build a feedback loop:
1. **Log conversations**: track questions and answers to identify hallucination patterns
2. **Test edge cases**: regularly test with questions that should be refused or answered with uncertainty
3. **Refine prompts**: update your system prompt based on observed failures
4. **Review source documents**: sometimes hallucination occurs because the indexed data itself is ambiguous or incomplete. Improving document quality is often the most effective fix
## Next steps
Set up scope restrictions and data grounding rules.
Let users verify AI responses by showing source documents.
Customize your workspace settings and system prompt.