> ## 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.
# Optimize chat prompts
> Improve conversational search response quality by tuning system prompts, document templates, and index chat settings.
The quality of conversational search responses depends on three layers of configuration: the system prompt, the document template, and the index-level chat settings. Each layer shapes what the LLM receives and how it responds. This guide covers how to tune each one for better results.
## System prompt strategies
The system prompt (set through [workspace settings](/capabilities/conversational_search/how_to/configure_chat_workspace)) defines the LLM's overall behavior. Beyond basic [guardrails](/capabilities/conversational_search/how_to/configure_guardrails), you can shape response quality with specific instructions.
### Be specific about the domain
Generic prompts produce generic answers. Tell the LLM exactly what it is and what data it works with:
```text Bad theme={null}
You are a helpful assistant.
```
```text Good theme={null}
You are a product specialist for an outdoor equipment store.
You help customers find hiking, camping, and climbing gear
based on their needs, experience level, and budget.
The search results contain our current product catalog
with prices, specifications, and customer reviews.
```
The more context the LLM has about the domain, the better it can interpret ambiguous queries and structure relevant answers.
### Define answer structure
Tell the LLM how to format responses. This improves consistency and readability:
```text System prompt theme={null}
When recommending products:
1. Start with a brief answer to the user's question
2. List 2-3 recommended products with their key specs
3. Explain why each product fits the user's needs
4. Mention the price range
When comparing products:
1. Create a brief comparison of the key differences
2. Recommend which product fits best based on the user's stated needs
3. Mention any trade-offs
```
### Control response length
Without guidance, LLMs tend to produce long responses. Set explicit length expectations:
```text System prompt theme={null}
Keep responses concise. For simple factual questions, answer
in 1-2 sentences. For product recommendations, use 3-5 short
paragraphs. For comparisons, use a brief list format.
Never exceed 300 words unless the user explicitly asks for
a detailed explanation.
```
## Tune the tool prompts
Beyond `prompts.system`, the workspace `prompts` object exposes three tool-facing prompts that Meilisearch injects into the function-calling schema the agent sees. They do not talk to the end user: they talk to the LLM about how to drive the Meilisearch search tool. Small edits here shift which index the agent picks, how it rewrites the query, and whether it decides to search at all.
| Field | What it configures |
| ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `prompts.searchDescription` | Internal description of the Meilisearch chat tools. Use it to instruct the agent on how and when to use the configured tools, for example encouraging it to search for any factual question or to skip the search for greetings. |
| `prompts.searchQParam` | Describes the expected user input and the desired query shape. Use it to tell the agent how to reformulate user messages into effective search queries and what output to expect back. |
| `prompts.searchIndexUidParam` | Describes each index the agent has access to and explains how to pick between them. This is the main lever when you have multiple indexes and want the agent to route queries correctly. |
```bash cURL theme={null}
curl \
-X PATCH 'MEILISEARCH_URL/chats/WORKSPACE_NAME/settings' \
-H 'Authorization: Bearer MEILISEARCH_KEY' \
-H 'Content-Type: application/json' \
--data-binary '{
"prompts": {
"searchDescription": "Search the outdoor gear catalog whenever the user asks about products, prices, availability, or specifications. Do not use it for generic small talk.",
"searchQParam": "A short natural-language query capturing the user intent. Rewrite long questions into 3 to 8 keywords. Preserve product names and brand names verbatim.",
"searchIndexUidParam": "Choose between: \"products\" for physical gear, accessories, and apparel; \"guides\" for how-to articles and buying guides; \"reviews\" for customer reviews and ratings. Use \"products\" by default."
}
}'
```
Change one prompt at a time and re-run your evaluation queries. Tool prompts compound: a change in `searchIndexUidParam` can make earlier `searchQParam` wording look wrong even though it hasn't changed.
## Configure index chat settings
Each index has chat-specific settings that control how documents are prepared for the LLM. Configure these through the [index chat settings](/capabilities/conversational_search/how_to/configure_index_chat_settings) endpoint:
```bash theme={null}
curl \
-X PATCH 'MEILISEARCH_URL/chats/WORKSPACE_NAME/indexes/products' \
-H 'Authorization: Bearer MEILISEARCH_KEY' \
-H 'Content-Type: application/json' \
--data-binary '{
"description": "Product catalog with hiking, camping, and climbing equipment. Each product has a name, description, price, category, brand, weight, and customer rating.",
"searchParameters": {
"limit": 5,
"hybrid": {
"semanticRatio": 0.7,
"embedder": "my-embedder"
},
"attributesToRetrieve": ["name", "description", "price", "category", "rating"]
}
}'
```
### Write a good index description
The `description` field tells the LLM what kind of data the index contains. The LLM uses this to decide whether to search the index and how to interpret results:
```text Bad theme={null}
Products index.
```
```text Good theme={null}
Product catalog for an outdoor equipment retailer. Contains
hiking boots, backpacks, tents, climbing gear, and camping
accessories. Each product includes name, detailed description,
price in USD, weight in grams, brand, category, average
customer rating (1-5), and number of reviews.
```
### Limit retrieved attributes
By default, Meilisearch sends all document attributes to the LLM. This can include irrelevant data that confuses the model or wastes tokens. Use `attributesToRetrieve` to send only what matters:
```json theme={null}
{
"searchParameters": {
"attributesToRetrieve": ["name", "description", "price", "rating"]
}
}
```
Exclude internal IDs, timestamps, image URLs, and other fields the LLM does not need for generating answers.
### Tune search parameters for chat
Conversational queries are often longer and more natural than keyword searches. Adjust search parameters to match:
* **Higher `semanticRatio`** (0.6-0.8): natural language questions benefit from semantic search more than keyword matching
* **Lower `limit`** (3-5): the LLM processes fewer, more relevant documents better than many loosely related ones
* **Broader matching strategy**: use `"matchingStrategy": "last"` (the default) to match as many terms as possible
```json theme={null}
{
"searchParameters": {
"limit": 5,
"hybrid": {
"semanticRatio": 0.7,
"embedder": "my-embedder"
},
"matchingStrategy": "last"
}
}
```
## Optimize document templates for chat
If your index uses an [embedder](/capabilities/hybrid_search/how_to/choose_an_embedder), the `documentTemplate` affects both embedding quality and the text the LLM sees during conversational search. A good template for chat should be readable as natural language:
```text Bad theme={null}
{{doc.name}} {{doc.price}} {{doc.category}}
```
```text Good theme={null}
{{doc.name}} is a {{doc.category}} product priced at ${{doc.price}}.
{{doc.description}}. Rated {{doc.rating}} out of 5 by customers.
```
The LLM reads these rendered templates as context. Structured, readable text helps it generate better answers. See [document template best practices](/capabilities/hybrid_search/advanced/document_template_best_practices) for detailed guidance.
## Test and iterate
After configuring prompts and settings, test with realistic queries to evaluate quality:
1. **Factual questions**: "What is the lightest 2-person tent you carry?" (should cite specific products with weights)
2. **Comparison questions**: "Should I get the TrailRunner Pro or the SpeedHike 3?" (should compare features)
3. **Vague questions**: "I need something for a weekend trip" (should ask clarifying questions or give broad recommendations)
4. **Out-of-scope questions**: "What is the weather forecast?" (should decline politely)
For each test, evaluate:
* Is the answer grounded in the search results?
* Is the response length appropriate?
* Does the formatting match your instructions?
* Are the recommended documents relevant to the question?
Adjust the system prompt, index description, and search parameters based on what you find.
## Next steps
Restrict responses to indexed data and defined topics
Full reference for index-level chat configuration
Write effective templates for embedding and chat