Merge pull request #3531 from meilisearch/fix-conversational-search-accuracy

Improve conversational search documentation

Author: qdequele

capabilities/conversational_search/advanced/chat_tooling_reference.mdx

@@ -5,6 +5,10 @@ description: An exhaustive reference of special chat tools supported by Meilisea
 
 When creating your conversational search agent, you may be able to extend the model's capabilities with a number of tools. This page lists Meilisearch-specific tools that may improve user experience.
 
+<Note>
+In code examples, replace `WORKSPACE_NAME` with the name of your workspace. On Meilisearch Cloud, the default workspace name is `cloud`.
+</Note>
+
 ## Meilisearch chat tools
 
 For the best user experience, configure all following tools.
@@ -112,6 +116,7 @@ curl \
   -H 'Content-Type: application/json' \
   --data-binary '{
     "model": "PROVIDER_MODEL_UID",
+    "stream": true,
     "messages": [
       {
         "role": "user",
@@ -145,7 +150,7 @@ curl \
           "parameters": {
             "type": "object",
             "properties": {
-              "role": { "type": "string", "description": "The role of the messages author, either user or assistant" },
+              "role": { "type": "string", "description": "The role of the messages author, either `tool` or `assistant`" },
               "content": { "type": "string", "description": "The contents of the assistant or tool message. Required unless tool_calls is specified." },
               "tool_calls": {
                 "type": ["array", "null"],
@@ -183,7 +188,7 @@ curl \
             "type": "object",
             "properties": {
               "call_id": { "type": "string", "description": "The call ID to track the original search associated to those sources" },
-              "documents": { "type": "object", "description": "The documents associated with the search. Only displayed attributes are returned" }
+              "documents": { "type": "array", "items": { "type": "object" }, "description": "The documents associated with the search. Only displayed attributes are returned" }
             },
             "required": ["call_id", "documents"],
             "additionalProperties": false
@@ -223,7 +228,7 @@ const tools = [
       parameters: {
         type: 'object',
         properties: {
-          role: { type: 'string', description: 'The role of the messages author, either user or assistant' },
+          role: { type: 'string', description: 'The role of the messages author, either `tool` or `assistant`' },
           content: { type: 'string', description: 'The contents of the assistant or tool message.' },
           tool_calls: { type: ['array', 'null'], description: 'The tool calls generated by the model', items: { type: 'object', properties: { function: { type: 'object', properties: { name: { type: 'string' }, arguments: { type: 'string' } } }, id: { type: 'string' }, type: { type: 'string' } } } },
           tool_call_id: { type: ['string', 'null'], description: 'Tool call that this message is responding to' },
@@ -243,7 +248,7 @@ const tools = [
         type: 'object',
         properties: {
           call_id: { type: 'string', description: 'The call ID to track the original search associated to those sources' },
-          documents: { type: 'object', description: 'The documents associated with the search.' },
+          documents: { type: 'array', items: { type: 'object' }, description: 'The documents associated with the search.' },
         },
         required: ['call_id', 'documents'],
         additionalProperties: false,
@@ -263,6 +268,7 @@ const response = await fetch(
     },
     body: JSON.stringify({
       model: 'PROVIDER_MODEL_UID',
+      stream: true,
       messages: [{ role: 'user', content: 'What are the best sci-fi movies?' }],
       tools,
     }),
@@ -342,7 +348,7 @@ const tools = [
         type: 'object',
         properties: {
           call_id: { type: 'string' },
-          documents: { type: 'object' },
+          documents: { type: 'array', items: { type: 'object' } },
         },
         required: ['call_id', 'documents'],
         additionalProperties: false,
@@ -422,7 +428,7 @@ const { textStream } = streamText({
         type: 'object',
         properties: {
           call_id: { type: 'string' },
-          documents: { type: 'object' },
+          documents: { type: 'array', items: { type: 'object' } },
         },
         required: ['call_id', 'documents'],
         additionalProperties: false,

capabilities/conversational_search/advanced/reduce_hallucination.mdx

@@ -14,7 +14,7 @@ When Meilisearch sends retrieved documents to the LLM, the model may:
 - 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, never its general knowledge**. All the techniques below reinforce this principle.
+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