Add support for multi-modal search

.code-samples.meilisearch.yaml

@@ -767,6 +767,20 @@ get_batch_1: |-
   client.get_batch(BATCH_UID)
 get_similar_post_1: |-
   client.index("INDEX_NAME").get_similar_documents({"id": "TARGET_DOCUMENT_ID", "embedder": "default"})
+search_parameter_reference_media_1: |-
+  client.index('movies_fragments').search(
+      "",
+      {
+          "media": {
+              "image_url": "https://image.tmdb.org/t/p/w500/gEU2QniE6E77NI6lCU6MxlNBvIx.jpg"
+          },
+          "hybrid": {
+              "embedder": "voyage",
+              "semanticRatio": 1.0
+          },
+          "limit": 3
+      }
+  )
 webhooks_get_1: |-
   client.get_webhooks()
 webhooks_get_single_1: |-
@@ -783,3 +797,4 @@ webhooks_patch_1: |-
   })
 webhooks_delete_1: |-
   client.delete_webhook('WEBHOOK_UID')
+

docs/conf.py

@@ -58,9 +58,9 @@
 # This value contains a list of modules to be mocked up.
 autodoc_mock_imports = ["camel_converter"]
 
-html_title = 'Meilisearch Python | Documentation'
+html_title = "Meilisearch Python | Documentation"
 
 # Add Fathom analytics script
 html_js_files = [
-    ("https://cdn.usefathom.com/script.js", { "data-site": "QNBPJXIV", "defer": "defer" })
+    ("https://cdn.usefathom.com/script.js", {"data-site": "QNBPJXIV", "defer": "defer"})
 ]

meilisearch/client.py

@@ -1110,3 +1110,23 @@ def _valid_uuid(uuid: str) -> bool:
         )
         match = uuid4hex.match(uuid)
         return bool(match)
+
+    def get_experimental_features(self) -> dict:
+        """
+        Retrieve the current settings for all experimental features.
+        Returns:
+            dict: A mapping of feature names to their enabled/disabled state.
+        """
+        return self.http.get(self.config.paths.experimental_features)
+
+    def update_experimental_features(self, features: dict) -> dict:
+        """
+        Update one or more experimental features.
+
+        Args:
+            features (dict): A dictionary mapping feature names to booleans.
+                             For example, {"multimodal": True} to enable multimodal.
+        Returns:
+            dict: The updated experimental features settings.
+        """
+        return self.http.patch(self.config.paths.experimental_features, body=features)

meilisearch/config.py

@@ -47,6 +47,7 @@ class Paths:
         localized_attributes = "localized-attributes"
         edit = "edit"
         network = "network"
+        experimental_features = "experimental-features"
         webhooks = "webhooks"
 
     def __init__(

meilisearch/models/embedders.py

@@ -167,6 +167,10 @@ class RestEmbedder(CamelBase):
         Template defining the data Meilisearch sends to the embedder
     document_template_max_bytes: Optional[int]
         Maximum allowed size of rendered document template (defaults to 400)
+    indexing_fragments: Optional[Dict[str, Dict[str, str]]]
+        Defines how to fragment documents for indexing (multi-modal search)
+    search_fragments: Optional[Dict[str, Dict[str, str]]]
+        Defines how to fragment search queries (multi-modal search)
     request: Dict[str, Any]
         A JSON value representing the request Meilisearch makes to the remote embedder
     response: Dict[str, Any]
@@ -185,6 +189,8 @@ class RestEmbedder(CamelBase):
     dimensions: Optional[int] = None
     document_template: Optional[str] = None
     document_template_max_bytes: Optional[int] = None
+    indexing_fragments: Optional[Dict[str, Dict[str, str]]] = None
+    search_fragments: Optional[Dict[str, Dict[str, str]]] = None
     request: Dict[str, Any]
     response: Dict[str, Any]
     headers: Optional[Dict[str, str]] = None

tests/client/test_experimental_features.py

@@ -0,0 +1,193 @@
+# pylint: disable=redefined-outer-name
+"""Tests for indexingFragments and searchFragments in embedders (multimodal feature).
+
+These tests validate CONFIGURATION ONLY, not AI functionality.
+They only ensure fragments can be configured and stored in Meilisearch.
+No AI calls or document indexing/searching occurs.
+"""
+
+import pytest
+
+DUMMY_URL = "http://localhost:8000/embed"
+TEST_MODEL = "test-model"
+MULTIMODAL_MODEL = "multimodal"
+
+
+def apply_embedders(index, config):
+    """Helper to update embedders and wait for task completion."""
+    response = index.update_embedders(config)
+    update = index.wait_for_task(response.task_uid)
+    assert update.status == "succeeded"
+    return index.get_embedders()
+
+
+def test_rest_embedder_with_fragments(empty_index, multimodal_enabled):
+    """Tests that REST embedder can be configured with indexingFragments and searchFragments."""
+    index = empty_index()
+
+    config = {
+        "rest_fragments": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "apiKey": "test-key",
+            "dimensions": 512,
+            "indexingFragments": {"text": {"value": "{{doc.title}} - {{doc.description}}"}},
+            "searchFragments": {"text": {"value": "{{fragment}}"}},
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+            "headers": {"Authorization": "Bearer test-key"},
+        }
+    }
+
+    embedders = apply_embedders(index, config)
+
+    e = embedders.embedders["rest_fragments"]
+    assert e.source == "rest"
+    assert e.url == DUMMY_URL
+    assert e.dimensions == 512
+    assert e.indexing_fragments is not None
+    assert e.search_fragments is not None
+
+
+def test_rest_embedder_with_multiple_fragments(empty_index, multimodal_enabled):
+    """Tests REST embedder with multiple fragment types."""
+    index = empty_index()
+
+    config = {
+        "multi_fragments": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 1024,
+            "indexingFragments": {
+                "text": {"value": "{{doc.title}}"},
+                "description": {"value": "{{doc.overview}}"}
+            },
+            "searchFragments": {
+                "text": {"value": "{{fragment}}"},
+                "description": {"value": "{{fragment}}"}
+            },
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    embedders = apply_embedders(index, config)
+
+    e = embedders.embedders["multi_fragments"]
+    assert e.source == "rest"
+    assert len(e.indexing_fragments) >= 1
+    assert len(e.search_fragments) >= 1
+
+
+def test_fragments_without_document_template(empty_index, multimodal_enabled):
+    """Tests fragments can be used without documentTemplate."""
+    index = empty_index()
+
+    config = {
+        "fragments_only": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 512,
+            "indexingFragments": {"text": {"value": "{{doc.content}}"}},
+            "searchFragments": {"text": {"value": "{{fragment}}"}},
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    embedders = apply_embedders(index, config)
+    e = embedders.embedders["fragments_only"]
+    assert e.document_template is None
+    assert e.indexing_fragments is not None
+    assert e.search_fragments is not None
+
+
+def test_fragments_require_multimodal_feature(empty_index):
+    """Tests fragments require multimodal feature enabled."""
+    index = empty_index()
+
+    config = {
+        "test": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 512,
+            "indexingFragments": {"text": {"value": "{{doc.title}}"}},
+            "searchFragments": {"text": {"value": "{{fragment}}"}},
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    # May succeed or fail depending on server config; both are acceptable
+    try:
+        embedders = apply_embedders(index, config)
+        assert embedders.embedders["test"].indexing_fragments is not None
+    except Exception:
+        pass
+
+
+def test_update_fragments_separately(empty_index, multimodal_enabled):
+    """Tests updating indexingFragments and searchFragments separately."""
+    index = empty_index()
+
+    initial_config = {
+        "updatable": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 512,
+            "indexingFragments": {"text": {"value": "{{doc.title}}"}},
+            "searchFragments": {"text": {"value": "{{fragment}}"}},
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    apply_embedders(index, initial_config)
+
+    updated_config = {
+        "updatable": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 512,
+            "indexingFragments": {"text": {"value": "{{doc.title}} - {{doc.description}}"}},
+            "searchFragments": {"text": {"value": "{{fragment}}"}},
+            "request": {"input": ["{{fragment}}"], "model": TEST_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    embedders = apply_embedders(index, updated_config)
+    assert embedders.embedders["updatable"].indexing_fragments is not None
+
+
+def test_profile_picture_and_title_fragments(empty_index, multimodal_enabled):
+    """Tests real-world use case: user profiles with picture and title."""
+    index = empty_index()
+
+    config = {
+        "user_profile": {
+            "source": "rest",
+            "url": DUMMY_URL,
+            "dimensions": 768,
+            "indexingFragments": {
+                "user_name": {"value": "{{doc.name}}"},
+                "avatar": {"value": "{{doc.profile_picture_url}}"},
+                "biography": {"value": "{{doc.bio}}"},
+            },
+            "searchFragments": {
+                "user_name": {"value": "{{fragment}}"},
+                "avatar": {"value": "{{fragment}}"},
+                "biography": {"value": "{{fragment}}"},
+            },
+            "request": {"input": ["{{fragment}}"], "model": MULTIMODAL_MODEL},
+            "response": {"data": [{"embedding": "{{embedding}}"}]},
+        }
+    }
+
+    embedders = apply_embedders(index, config)
+    e = embedders.embedders["user_profile"]
+
+    assert e.source == "rest"
+    expected_keys = {"user_name", "avatar", "biography"}
+    assert set(e.indexing_fragments.keys()) == expected_keys
+    assert set(e.search_fragments.keys()) == expected_keys