Multimodal generation - Vector Retrieval Service for Milvus

The AI_MULTI_MODAL_GENERATE function converts text, images, videos, or audio into text output. Use it to describe product materials, summarize videos, answer visual questions, or transcribe speech.

Command format

Multimodal generation supports two usage patterns:

REST API — Call the endpoint directly to generate text from your input. Use this approach for ad-hoc or batch generation tasks.
Collection Function — Define a TEXTTRANSFORM Function in your collection schema to generate text automatically when data is inserted. Use this approach to integrate generation into your data ingestion pipeline.

The REST API examples require curl and jq. The Python examples require pymilvus. Replace the endpoint URL and credentials in each example with your Milvus instance values.

REST API

POST /v2/vectordb/ai/multi_modal_generate
{
  "model_name": "<model_name>",
  "texts": ["<text_or_media_url>"],
  "params": {"media_type": "text | image | video | audio", "prompt": "<instruction>"}
}

Python (Collection Function)

schema = MilvusClient.create_schema(auto_id=True, enable_dynamic_field=False)
schema.add_field("id", DataType.INT64, is_primary=True)
schema.add_field("content", DataType.VARCHAR, max_length=4096)
schema.add_field("generated", DataType.VARCHAR, max_length=1024)
schema.add_field("dummy_vector", DataType.FLOAT_VECTOR, dim=2)
schema.add_function(
    Function(
        name="generate_text",
        function_type=texttransform_function_type(),
        input_field_names=["content"],
        output_field_names=["generated"],
        params={
            "provider": "aliyun_milvus",
            "model_name": "<model_name>",
            "task": "ai_multi_modal_generate",
            "media_type": "text",
            "prompt": "Answer concisely: ${content}",
            "temperature": "0",
        },
    )
)

Parameters

Parameter	Description
`model_name`	Required. For text, image, and video, use `qwen3.7-plus` (recommended), `qwen3.6-plus`, `qwen3.6-flash`, `qwen3.5-flash`, or `kimi-k2.6`. For audio, use `qwen3-asr-flash`.
`texts`	Required for REST API. For the `text` media type, pass the original text. For other media types, pass a model-accessible URL. Images can also be passed as Base64. Results match the input order.
`media_type`	Required. Valid values: `text`, `image`, `video`, `audio`.
`prompt`	Required for `image` and `video`. Optional for `text`. Not supported for `audio`. In a Collection Function schema, use `${field_name}` to reference field values.
`temperature`	Optional. Controls output generation behavior. Passed as a number in REST API calls or as a string in Collection Function params. Set to `0` in all examples.
`language` / `enable_itn`	Optional. Audio only. `language` specifies the recognition language (for example, `zh`, `en`). `enable_itn` controls whether to enable inverse text normalization.
`timeout_sec` / `max_concurrency`	Optional. Control the per-call timeout and batch concurrency.
`output_mapping`	Optional. Collection Function `text`, `image`, and `video` scenarios only. Maps simple paths in the model output JSON to multiple `VARCHAR`, `TEXT`, or `JSON` fields. Without this setting, only one text output field is allowed.
`provider` / `task`	Required for Collection Function only. Fixed values: `aliyun_milvus` and `ai_multi_modal_generate`.

Return values

On success, the response contains the following fields:

data.output.outputs — A text array. Each element corresponds one-to-one with the input texts array.
data.usage — Usage metrics that may include:
- input_tokens, output_tokens, total_tokens — Token counts.
- image_tokens, video_tokens, audio_tokens — Media-specific token counts, when applicable.
- seconds — Media duration, when applicable.

Example 1: Generate a concise text response (text)

This example generates a concise text response from a text input using the text media type.

REST API

#!/usr/bin/env bash
set -euo pipefail

MILVUS_REST_BASE_URL="http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_AUTH_TOKEN="<yourUsername>:<yourPassword>"

post_json() {
  local path="$1"
  local body="$2"
  curl -X POST \
    "$MILVUS_REST_BASE_URL$path" \
    -H "Authorization: Bearer $MILVUS_AUTH_TOKEN" \
    -H "Content-Type: application/json" \
    -d "$body"
}

BODY=$(cat <<JSON
{
  "model_name": "qwen3.7-plus",
  "texts": ["Explain in one sentence how Milvus is used in a RAG application."],
  "params": {"media_type": "text", "prompt": "Answer concisely.", "temperature": 0}
}
JSON
)
RESPONSE_BODY="$(post_json "/v2/vectordb/ai/multi_modal_generate" "$BODY")"
if command -v jq >/dev/null 2>&1; then
  echo "$RESPONSE_BODY" | jq .
  [ "$(echo "$RESPONSE_BODY" | jq -r '.code // -1')" = "0" ] || exit 1
else
  echo "$RESPONSE_BODY"
fi

Python

from __future__ import annotations

from typing import Any

from pymilvus import DataType, Function, FunctionType, MilvusClient

MILVUS_URI = "http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_TOKEN = "<yourUsername>:<yourPassword>"

DUMMY_VECTOR_DIM = 2
TEXTTRANSFORM_FUNCTION_TYPE = 9

def texttransform_function_type() -> Any:
    for type_name in ("TEXTTRANSFORM", "TEXT_TRANSFORM", "TextTransform"):
        function_type = getattr(FunctionType, type_name, None)
        if function_type is not None:
            return function_type
    # Alibaba Cloud Milvus provides TEXTTRANSFORM as a managed extension (function type value 9);
    # some pymilvus versions do not yet have this enum member built-in, and Function(...) validates via FunctionType(...).
    existing = getattr(FunctionType, "_value2member_map_", {}).get(TEXTTRANSFORM_FUNCTION_TYPE)
    if existing is not None:
        return existing
    extension = int.__new__(FunctionType, TEXTTRANSFORM_FUNCTION_TYPE)
    extension._name_ = "TEXTTRANSFORM"
    extension._value_ = TEXTTRANSFORM_FUNCTION_TYPE
    FunctionType._value2member_map_[TEXTTRANSFORM_FUNCTION_TYPE] = extension
    FunctionType._member_map_["TEXTTRANSFORM"] = extension
    return extension

def add_id(schema: Any) -> None:
    schema.add_field("id", DataType.INT64, is_primary=True)

def add_dummy_vector(schema: Any) -> None:
    schema.add_field("dummy_vector", DataType.FLOAT_VECTOR, dim=DUMMY_VECTOR_DIM)

def run_texttransform_example(*, client, collection_name, input_fields, output_field, function_name, function_params, rows) -> None:
    if client.has_collection(collection_name):
        client.drop_collection(collection_name)
    schema = MilvusClient.create_schema(auto_id=True, enable_dynamic_field=False)
    add_id(schema)
    for name, data_type, max_length in input_fields:
        field_params = {"max_length": max_length} if max_length is not None else {}
        schema.add_field(name, data_type, **field_params)
    output_name, output_data_type, output_max_length = output_field
    output_params = {"max_length": output_max_length} if output_max_length is not None else {}
    schema.add_field(output_name, output_data_type, **output_params)
    add_dummy_vector(schema)
    schema.add_function(
        Function(
            name=function_name,
            function_type=texttransform_function_type(),
            input_field_names=[name for name, _, _ in input_fields],
            output_field_names=[output_name],
            params=function_params,
        )
    )
    index_params = client.prepare_index_params()
    index_params.add_index(field_name="dummy_vector", index_type="AUTOINDEX", metric_type="COSINE")
    client.create_collection(collection_name=collection_name, schema=schema, index_params=index_params)
    client.insert(collection_name, rows)
    client.flush(collection_name)
    fields = [name for name, _, _ in input_fields] + [output_name]
    for row in client.query(collection_name, filter="", output_fields=fields, limit=len(rows)):
        print(row)

client = MilvusClient(uri=MILVUS_URI, token=MILVUS_TOKEN)

run_texttransform_example(
    client=client,
    collection_name="simple_ai_multi_modal_generate_text",
    input_fields=[("content", DataType.VARCHAR, 4096)],
    output_field=("generated", DataType.VARCHAR, 1024),
    function_name="generate_text",
    function_params={"provider": "aliyun_milvus", "model_name": "qwen3.7-plus", "task": "ai_multi_modal_generate", "media_type": "text", "prompt": "Answer concisely: ${content}", "temperature": "0"},
    rows=[{"content": "Explain how Milvus is used in a RAG application.", "dummy_vector": [0.1, 0.2]}],
)

Expected result: The generated field contains a concise response describing how Milvus is used in a RAG application. In the REST API response, this value is at data.output.outputs[0].

Example 2: Generate retrieval descriptions from product images (image)

This example generates an objective description for image-text retrieval before indexing clothing materials. The bash example requires jq.

REST API

#!/usr/bin/env bash
set -euo pipefail

MILVUS_REST_BASE_URL="http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_AUTH_TOKEN="<yourUsername>:<yourPassword>"

post_json() {
  local path="$1"
  local body="$2"
  curl -X POST \
    "$MILVUS_REST_BASE_URL$path" \
    -H "Authorization: Bearer $MILVUS_AUTH_TOKEN" \
    -H "Content-Type: application/json" \
    -d "$body"
}

BODY=$(cat <<JSON
{
  "model_name": "qwen3.7-plus",
  "texts": ["https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260415/hynnff/wan-video-edit-clothes.webp"],
  "params": {"media_type": "image", "prompt": "Describe the image in one concise sentence.", "temperature": 0}
}
JSON
)
RESPONSE_BODY="$(post_json "/v2/vectordb/ai/multi_modal_generate" "$BODY")"
if command -v jq >/dev/null 2>&1; then
  echo "$RESPONSE_BODY" | jq .
  [ "$(echo "$RESPONSE_BODY" | jq -r '.code // -1')" = "0" ] || exit 1
else
  echo "$RESPONSE_BODY"
fi

Python

from __future__ import annotations

from typing import Any

from pymilvus import DataType, Function, FunctionType, MilvusClient

MILVUS_URI = "http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_TOKEN = "<yourUsername>:<yourPassword>"

DUMMY_VECTOR_DIM = 2
TEXTTRANSFORM_FUNCTION_TYPE = 9

def texttransform_function_type() -> Any:
    for type_name in ("TEXTTRANSFORM", "TEXT_TRANSFORM", "TextTransform"):
        function_type = getattr(FunctionType, type_name, None)
        if function_type is not None:
            return function_type
    # Alibaba Cloud Milvus provides TEXTTRANSFORM as a managed extension (function type value 9);
    # some pymilvus versions do not yet have this enum member built-in, and Function(...) validates via FunctionType(...).
    existing = getattr(FunctionType, "_value2member_map_", {}).get(TEXTTRANSFORM_FUNCTION_TYPE)
    if existing is not None:
        return existing
    extension = int.__new__(FunctionType, TEXTTRANSFORM_FUNCTION_TYPE)
    extension._name_ = "TEXTTRANSFORM"
    extension._value_ = TEXTTRANSFORM_FUNCTION_TYPE
    FunctionType._value2member_map_[TEXTTRANSFORM_FUNCTION_TYPE] = extension
    FunctionType._member_map_["TEXTTRANSFORM"] = extension
    return extension

def add_id(schema: Any) -> None:
    schema.add_field("id", DataType.INT64, is_primary=True)

def add_dummy_vector(schema: Any) -> None:
    schema.add_field("dummy_vector", DataType.FLOAT_VECTOR, dim=DUMMY_VECTOR_DIM)

def run_texttransform_example(*, client, collection_name, input_fields, output_field, function_name, function_params, rows) -> None:
    if client.has_collection(collection_name):
        client.drop_collection(collection_name)
    schema = MilvusClient.create_schema(auto_id=True, enable_dynamic_field=False)
    add_id(schema)
    for name, data_type, max_length in input_fields:
        field_params = {"max_length": max_length} if max_length is not None else {}
        schema.add_field(name, data_type, **field_params)
    output_name, output_data_type, output_max_length = output_field
    output_params = {"max_length": output_max_length} if output_max_length is not None else {}
    schema.add_field(output_name, output_data_type, **output_params)
    add_dummy_vector(schema)
    schema.add_function(
        Function(
            name=function_name,
            function_type=texttransform_function_type(),
            input_field_names=[name for name, _, _ in input_fields],
            output_field_names=[output_name],
            params=function_params,
        )
    )
    index_params = client.prepare_index_params()
    index_params.add_index(field_name="dummy_vector", index_type="AUTOINDEX", metric_type="COSINE")
    client.create_collection(collection_name=collection_name, schema=schema, index_params=index_params)
    client.insert(collection_name, rows)
    client.flush(collection_name)
    fields = [name for name, _, _ in input_fields] + [output_name]
    for row in client.query(collection_name, filter="", output_fields=fields, limit=len(rows)):
        print(row)

client = MilvusClient(uri=MILVUS_URI, token=MILVUS_TOKEN)

run_texttransform_example(
    client=client,
    collection_name="simple_ai_multi_modal_generate_image",
    input_fields=[("image_url", DataType.VARCHAR, 4096)],
    output_field=("description", DataType.VARCHAR, 2048),
    function_name="describe_image",
    function_params={"provider": "aliyun_milvus", "model_name": "qwen3.7-plus", "task": "ai_multi_modal_generate", "media_type": "image", "prompt": "Describe the image in one concise sentence.", "temperature": "0"},
    rows=[{"image_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260415/hynnff/wan-video-edit-clothes.webp", "dummy_vector": [0.1, 0.2]}],
)

Expected result: The description field contains a concise objective description of the image, which can be written to a product retrieval field.

Example 3: Generate retrieval summaries from concept videos (video)

This example generates retrieval summaries from concept videos to enable searching by character, action, and visual atmosphere. Only validates that the text is non-empty; no fixed wording is assumed.

REST API

#!/usr/bin/env bash
set -euo pipefail

MILVUS_REST_BASE_URL="http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_AUTH_TOKEN="<yourUsername>:<yourPassword>"

post_json() {
  local path="$1"
  local body="$2"
  curl -X POST \
    "$MILVUS_REST_BASE_URL$path" \
    -H "Authorization: Bearer $MILVUS_AUTH_TOKEN" \
    -H "Content-Type: application/json" \
    -d "$body"
}

BODY=$(cat <<JSON
{
  "model_name": "qwen3.7-plus",
  "texts": ["https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260409/dozxak/Wan_Video_Edit_33_1.mp4"],
  "params": {"media_type": "video", "prompt": "Describe the video in one concise sentence.", "temperature": 0}
}
JSON
)
RESPONSE_BODY="$(post_json "/v2/vectordb/ai/multi_modal_generate" "$BODY")"
if command -v jq >/dev/null 2>&1; then
  echo "$RESPONSE_BODY" | jq .
  [ "$(echo "$RESPONSE_BODY" | jq -r '.code // -1')" = "0" ] || exit 1
else
  echo "$RESPONSE_BODY"
fi

Python

from __future__ import annotations

from typing import Any

from pymilvus import DataType, Function, FunctionType, MilvusClient

MILVUS_URI = "http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_TOKEN = "<yourUsername>:<yourPassword>"

DUMMY_VECTOR_DIM = 2
TEXTTRANSFORM_FUNCTION_TYPE = 9

def texttransform_function_type() -> Any:
    for type_name in ("TEXTTRANSFORM", "TEXT_TRANSFORM", "TextTransform"):
        function_type = getattr(FunctionType, type_name, None)
        if function_type is not None:
            return function_type
    # Alibaba Cloud Milvus provides TEXTTRANSFORM as a managed extension (function type value 9);
    # some pymilvus versions do not yet have this enum member built-in, and Function(...) validates via FunctionType(...).
    existing = getattr(FunctionType, "_value2member_map_", {}).get(TEXTTRANSFORM_FUNCTION_TYPE)
    if existing is not None:
        return existing
    extension = int.__new__(FunctionType, TEXTTRANSFORM_FUNCTION_TYPE)
    extension._name_ = "TEXTTRANSFORM"
    extension._value_ = TEXTTRANSFORM_FUNCTION_TYPE
    FunctionType._value2member_map_[TEXTTRANSFORM_FUNCTION_TYPE] = extension
    FunctionType._member_map_["TEXTTRANSFORM"] = extension
    return extension

def add_id(schema: Any) -> None:
    schema.add_field("id", DataType.INT64, is_primary=True)

def add_dummy_vector(schema: Any) -> None:
    schema.add_field("dummy_vector", DataType.FLOAT_VECTOR, dim=DUMMY_VECTOR_DIM)

def run_texttransform_example(*, client, collection_name, input_fields, output_field, function_name, function_params, rows) -> None:
    if client.has_collection(collection_name):
        client.drop_collection(collection_name)
    schema = MilvusClient.create_schema(auto_id=True, enable_dynamic_field=False)
    add_id(schema)
    for name, data_type, max_length in input_fields:
        field_params = {"max_length": max_length} if max_length is not None else {}
        schema.add_field(name, data_type, **field_params)
    output_name, output_data_type, output_max_length = output_field
    output_params = {"max_length": output_max_length} if output_max_length is not None else {}
    schema.add_field(output_name, output_data_type, **output_params)
    add_dummy_vector(schema)
    schema.add_function(
        Function(
            name=function_name,
            function_type=texttransform_function_type(),
            input_field_names=[name for name, _, _ in input_fields],
            output_field_names=[output_name],
            params=function_params,
        )
    )
    index_params = client.prepare_index_params()
    index_params.add_index(field_name="dummy_vector", index_type="AUTOINDEX", metric_type="COSINE")
    client.create_collection(collection_name=collection_name, schema=schema, index_params=index_params)
    client.insert(collection_name, rows)
    client.flush(collection_name)
    fields = [name for name, _, _ in input_fields] + [output_name]
    for row in client.query(collection_name, filter="", output_fields=fields, limit=len(rows)):
        print(row)

client = MilvusClient(uri=MILVUS_URI, token=MILVUS_TOKEN)

run_texttransform_example(
    client=client,
    collection_name="simple_ai_multi_modal_generate_video",
    input_fields=[("video_url", DataType.VARCHAR, 4096)],
    output_field=("description", DataType.VARCHAR, 2048),
    function_name="describe_video",
    function_params={"provider": "aliyun_milvus", "model_name": "qwen3.7-plus", "task": "ai_multi_modal_generate", "media_type": "video", "prompt": "Describe the video in one concise sentence.", "temperature": "0"},
    rows=[{"video_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260409/dozxak/Wan_Video_Edit_33_1.mp4", "dummy_vector": [0.1, 0.2]}],
)

Expected result: The description field contains a one-sentence summary of the video content, which can be used as a creative material retrieval field.

Example 4: Transcribe and archive a hotline greeting (audio)

This example transcribes a hotline greeting into searchable text for archiving. Audio requests do not require a prompt. The bash example requires jq. The sample audio is a public greeting; the transcription language depends on the actual audio content.

REST API

#!/usr/bin/env bash
set -euo pipefail

MILVUS_REST_BASE_URL="http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_AUTH_TOKEN="<yourUsername>:<yourPassword>"

post_json() {
  local path="$1"
  local body="$2"
  curl -X POST \
    "$MILVUS_REST_BASE_URL$path" \
    -H "Authorization: Bearer $MILVUS_AUTH_TOKEN" \
    -H "Content-Type: application/json" \
    -d "$body"
}

BODY=$(cat <<JSON
{
  "model_name": "qwen3-asr-flash",
  "texts": ["https://dashscope.oss-cn-beijing.aliyuncs.com/audios/welcome.mp3"],
  "params": {"media_type": "audio", "language": "en", "enable_itn": true}
}
JSON
)
RESPONSE_BODY="$(post_json "/v2/vectordb/ai/multi_modal_generate" "$BODY")"
if command -v jq >/dev/null 2>&1; then
  echo "$RESPONSE_BODY" | jq .
  [ "$(echo "$RESPONSE_BODY" | jq -r '.code // -1')" = "0" ] || exit 1
else
  echo "$RESPONSE_BODY"
fi

Python

from __future__ import annotations

from typing import Any

from pymilvus import DataType, Function, FunctionType, MilvusClient

MILVUS_URI = "http://c-xxxx.milvus.aliyuncs.com:19530"
MILVUS_TOKEN = "<yourUsername>:<yourPassword>"

DUMMY_VECTOR_DIM = 2
TEXTTRANSFORM_FUNCTION_TYPE = 9

def texttransform_function_type() -> Any:
    for type_name in ("TEXTTRANSFORM", "TEXT_TRANSFORM", "TextTransform"):
        function_type = getattr(FunctionType, type_name, None)
        if function_type is not None:
            return function_type
    # Alibaba Cloud Milvus provides TEXTTRANSFORM as a managed extension (function type value 9);
    # some pymilvus versions do not yet have this enum member built-in, and Function(...) validates via FunctionType(...).
    existing = getattr(FunctionType, "_value2member_map_", {}).get(TEXTTRANSFORM_FUNCTION_TYPE)
    if existing is not None:
        return existing
    extension = int.__new__(FunctionType, TEXTTRANSFORM_FUNCTION_TYPE)
    extension._name_ = "TEXTTRANSFORM"
    extension._value_ = TEXTTRANSFORM_FUNCTION_TYPE
    FunctionType._value2member_map_[TEXTTRANSFORM_FUNCTION_TYPE] = extension
    FunctionType._member_map_["TEXTTRANSFORM"] = extension
    return extension

def add_id(schema: Any) -> None:
    schema.add_field("id", DataType.INT64, is_primary=True)

def add_dummy_vector(schema: Any) -> None:
    schema.add_field("dummy_vector", DataType.FLOAT_VECTOR, dim=DUMMY_VECTOR_DIM)

def run_texttransform_example(*, client, collection_name, input_fields, output_field, function_name, function_params, rows) -> None:
    if client.has_collection(collection_name):
        client.drop_collection(collection_name)
    schema = MilvusClient.create_schema(auto_id=True, enable_dynamic_field=False)
    add_id(schema)
    for name, data_type, max_length in input_fields:
        field_params = {"max_length": max_length} if max_length is not None else {}
        schema.add_field(name, data_type, **field_params)
    output_name, output_data_type, output_max_length = output_field
    output_params = {"max_length": output_max_length} if output_max_length is not None else {}
    schema.add_field(output_name, output_data_type, **output_params)
    add_dummy_vector(schema)
    schema.add_function(
        Function(
            name=function_name,
            function_type=texttransform_function_type(),
            input_field_names=[name for name, _, _ in input_fields],
            output_field_names=[output_name],
            params=function_params,
        )
    )
    index_params = client.prepare_index_params()
    index_params.add_index(field_name="dummy_vector", index_type="AUTOINDEX", metric_type="COSINE")
    client.create_collection(collection_name=collection_name, schema=schema, index_params=index_params)
    client.insert(collection_name, rows)
    client.flush(collection_name)
    fields = [name for name, _, _ in input_fields] + [output_name]
    for row in client.query(collection_name, filter="", output_fields=fields, limit=len(rows)):
        print(row)

client = MilvusClient(uri=MILVUS_URI, token=MILVUS_TOKEN)

run_texttransform_example(
    client=client,
    collection_name="simple_ai_multi_modal_generate_audio",
    input_fields=[("audio_url", DataType.VARCHAR, 4096)],
    output_field=("transcript", DataType.VARCHAR, 4096),
    function_name="transcribe_audio",
    function_params={"provider": "aliyun_milvus", "model_name": "qwen3-asr-flash", "task": "ai_multi_modal_generate", "media_type": "audio", "language": "en", "enable_itn": "true"},
    rows=[{"audio_url": "https://dashscope.oss-cn-beijing.aliyuncs.com/audios/welcome.mp3", "dummy_vector": [0.1, 0.2]}],
)

Expected result: The transcript field contains non-empty transcription text. The sample audio transcribes to Welcome to Alibaba Cloud., rendered in English; the API returns the transcription in the source language of the audio. The transcription can be archived, but do not use the example code to automatically draw compliance conclusions.

Media security and retention

Before using your own images, videos, or audio, confirm that you have obtained the necessary permissions for the content, portrait rights, and voice rights. Use least-privilege URLs with expiration, and delete personal information in transcriptions and derived text in accordance with your business retention policy.