be/src/exprs/function/ai/embed.h

Source
// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#pragma once

#include <glog/logging.h>
#include <rapidjson/document.h>

#include <string_view>

#include "core/data_type/data_type_nullable.h"
#include "core/data_type/primitive_type.h"
#include "exprs/function/ai/ai_functions.h"
#include "util/jsonb_utils.h"
#include "util/s3_uri.h"
#include "util/s3_util.h"

namespace doris {
class FunctionEmbed : public AIFunction<FunctionEmbed> {
public:
    static constexpr auto name = "embed";

    static constexpr size_t number_of_arguments = 2;

    static constexpr auto system_prompt = "";

    DataTypePtr get_nested_return_type_impl(const DataTypes& /*arguments*/) const {
        return std::make_shared<DataTypeArray>(make_nullable(std::make_shared<DataTypeFloat32>()));
    }

    using PreparedFunctionImpl::execute;

    AIResource select_ai_resource(const TAIResource& resource, PrimitiveType input_type) const {
        bool has_complete_multimodal_embed_properties = _has_complete_resource_properties(
                resource.embed_mm_endpoint, resource.embed_mm_provider_type,
                resource.embed_mm_model_name, resource.embed_mm_api_key);
        if (input_type == PrimitiveType::TYPE_JSONB && has_complete_multimodal_embed_properties) {
            return AIResource::from_multimodal_embed(resource);
        }
        bool has_complete_embed_properties = _has_complete_resource_properties(
                resource.embed_endpoint, resource.embed_provider_type, resource.embed_model_name,
                resource.embed_api_key);
        return has_complete_embed_properties ? AIResource::from_embed(resource)
                                             : AIResource(resource);
    }

    Status execute(FunctionContext* context, Block& block, const ColumnNumbers& arguments,
                   uint32_t result, size_t input_rows_count, const AIResource& config,
                   std::shared_ptr<AIAdapter>& adapter) const {
        if (arguments.size() != 2) {
            return Status::InvalidArgument("Function EMBED expects 2 arguments, but got {}",
                                           arguments.size());
        }

        const auto& input = block.get_by_position(arguments[1]);
        ColumnUInt8::MutablePtr result_null_map;
        if (input.type->is_nullable()) {
            const auto& [column, is_const] = unpack_if_const(input.column);
            const auto& nullable =
                    assert_cast<const ColumnNullable&, TypeCheckOnRelease::DISABLE>(*column);
            result_null_map = ColumnUInt8::create(input_rows_count, 0);
            VectorizedUtils::update_null_map(result_null_map->get_data(),
                                             nullable.get_null_map_data(), is_const);
        }

        if (result_null_map &&
            !simd::contain_zero(result_null_map->get_data().data(), input_rows_count)) {
            block.get_by_position(result).column =
                    block.get_by_position(result).type->create_column_const(input_rows_count,
                                                                            Field());
            return Status::OK();
        }

        ColumnPtr input_column =
                input.unnest_nullable(input.type->is_nullable() ? input.get_nullable_column_info()
                                                                : NullableColumnInfo {},
                                      false)
                        .column;
        PrimitiveType input_type = remove_nullable(input.type)->get_primitive_type();
        if (input_type == PrimitiveType::TYPE_JSONB) {
            return _execute_multimodal_embed(context, block, result, input_rows_count, config,
                                             adapter, input_column, std::move(result_null_map));
        }
        if (input_type == PrimitiveType::TYPE_STRING || input_type == PrimitiveType::TYPE_VARCHAR ||
            input_type == PrimitiveType::TYPE_CHAR) {
            return _execute_text_embed(context, block, result, input_rows_count, config, adapter,
                                       input_column, std::move(result_null_map));
        }
        return Status::InvalidArgument(
                "Function EMBED expects the second argument to be STRING or JSON, but got type {}",
                input.type->get_name());
    }

    static FunctionPtr create() { return std::make_shared<FunctionEmbed>(); }

private:
    static bool _has_complete_resource_properties(std::string_view endpoint,
                                                  std::string_view provider_type,
                                                  std::string_view model_name,
                                                  std::string_view api_key) {
        return !endpoint.empty() && !provider_type.empty() && !model_name.empty() &&
               (provider_type == "LOCAL" || !api_key.empty());
    }

    static int32_t _get_embed_max_batch_size(FunctionContext* context) {
        QueryContext* query_ctx = context->state()->get_query_ctx();
        DORIS_CHECK(query_ctx != nullptr);

        return query_ctx->query_options().embed_max_batch_size;
    }

    Status _execute_text_embed(FunctionContext* context, Block& block, uint32_t result,
                               size_t input_rows_count, const AIResource& config,
                               std::shared_ptr<AIAdapter>& adapter, const ColumnPtr& input_column,
                               ColumnUInt8::MutablePtr result_null_map) const {
        auto col_result = ColumnArray::create(
                ColumnNullable::create(ColumnFloat32::create(), ColumnUInt8::create()));
        std::vector<std::string> batch_prompts;
        size_t current_batch_size = 0;
        const int32_t max_batch_size = _get_embed_max_batch_size(context);
        const size_t max_context_window_size =
                static_cast<size_t>(get_ai_context_window_size(context));
        const NullMap* null_map = result_null_map ? &result_null_map->get_data() : nullptr;
        const Columns prompt_columns {input_column};

        for (size_t i = 0; i < input_rows_count; ++i) {
            if (null_map && (*null_map)[i]) {
                continue;
            }

            std::string prompt;
            RETURN_IF_ERROR(build_prompt(prompt_columns, i, prompt));

            const size_t prompt_size = prompt.size();

            if (prompt_size > max_context_window_size) {
                // flush history batch
                RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
                                                            adapter, context));
                current_batch_size = 0;

                batch_prompts.emplace_back(std::move(prompt));
                RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
                                                            adapter, context));
                continue;
            }

            if (!batch_prompts.empty() &&
                (current_batch_size + prompt_size > max_context_window_size ||
                 batch_prompts.size() >= static_cast<size_t>(max_batch_size))) {
                RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
                                                            adapter, context));
                current_batch_size = 0;
            }

            batch_prompts.emplace_back(std::move(prompt));
            current_batch_size += prompt_size;
        }

        RETURN_IF_ERROR(
                _flush_text_embedding_batch(batch_prompts, *col_result, config, adapter, context));

        block.replace_by_position(result, _expand_and_wrap_nullable_result(
                                                  std::move(col_result), std::move(result_null_map),
                                                  input_rows_count));
        return Status::OK();
    }

    Status _execute_multimodal_embed(FunctionContext* context, Block& block, uint32_t result,
                                     size_t input_rows_count, const AIResource& config,
                                     std::shared_ptr<AIAdapter>& adapter,
                                     const ColumnPtr& input_column,
                                     ColumnUInt8::MutablePtr result_null_map) const {
        auto col_result = ColumnArray::create(
                ColumnNullable::create(ColumnFloat32::create(), ColumnUInt8::create()));
        std::vector<MultimodalType> batch_media_types;
        std::vector<std::string> batch_media_content_types;
        std::vector<std::string> batch_media_urls;
        const NullMap* null_map = result_null_map ? &result_null_map->get_data() : nullptr;

        int64_t ttl_seconds = 3600;
        QueryContext* query_ctx = context->state()->get_query_ctx();
        if (query_ctx && query_ctx->query_options().__isset.file_presigned_url_ttl_seconds) {
            ttl_seconds = query_ctx->query_options().file_presigned_url_ttl_seconds;
            if (ttl_seconds <= 0) {
                ttl_seconds = 3600;
            }
        }

        const int32_t max_batch_size = _get_embed_max_batch_size(context);

        for (size_t i = 0; i < input_rows_count; ++i) {
            if (null_map && (*null_map)[i]) {
                continue;
            }

            rapidjson::Document file_input;
            RETURN_IF_ERROR(_parse_file_input(*input_column, i, file_input));

            std::string content_type;
            MultimodalType media_type;
            RETURN_IF_ERROR(_infer_media_type(file_input, content_type, media_type));

            std::string media_url;
            RETURN_IF_ERROR(_resolve_media_url(file_input, ttl_seconds, media_url));

            if (!batch_media_urls.empty() &&
                batch_media_urls.size() >= static_cast<size_t>(max_batch_size)) {
                RETURN_IF_ERROR(_flush_multimodal_embedding_batch(
                        batch_media_types, batch_media_content_types, batch_media_urls, *col_result,
                        config, adapter, context));
            }

            batch_media_types.emplace_back(media_type);
            batch_media_content_types.emplace_back(std::move(content_type));
            batch_media_urls.emplace_back(std::move(media_url));
        }

        RETURN_IF_ERROR(_flush_multimodal_embedding_batch(
                batch_media_types, batch_media_content_types, batch_media_urls, *col_result, config,
                adapter, context));

        block.replace_by_position(result, _expand_and_wrap_nullable_result(
                                                  std::move(col_result), std::move(result_null_map),
                                                  input_rows_count));
        return Status::OK();
    }

    // EMBED-private helper.
    // Sends one embedding request with a prebuilt request body and validates returned row count.
    Status _execute_prebuilt_embedding_request(const std::string& request_body,
                                               std::vector<std::vector<float>>& results,
                                               size_t expected_size, const AIResource& config,
                                               std::shared_ptr<AIAdapter>& adapter,
                                               FunctionContext* context) const {
        std::string response;
#ifdef BE_TEST
        if (config.provider_type == "MOCK") {
            results.clear();
            results.reserve(expected_size);
            for (size_t i = 0; i < expected_size; ++i) {
                results.emplace_back(std::initializer_list<float> {0, 1, 2, 3, 4});
            }
            return Status::OK();
        }
#endif

        RETURN_IF_ERROR(
                this->send_request_to_llm(request_body, response, config, adapter, context));

        RETURN_IF_ERROR(adapter->parse_embedding_response(response, results));
        if (results.empty()) {
            return Status::InternalError("AI returned empty result");
        }
        if (results.size() != expected_size) [[unlikely]] {
            return Status::InternalError(
                    "AI embedding returned {} results, but {} inputs were sent", results.size(),
                    expected_size);
        }
        return Status::OK();
    }

    // EMBED-private helper.
    // Flushes one accumulated text embedding batch into the output array column.
    Status _flush_text_embedding_batch(std::vector<std::string>& batch_prompts,
                                       ColumnArray& col_result, const AIResource& config,
                                       std::shared_ptr<AIAdapter>& adapter,
                                       FunctionContext* context) const {
        if (batch_prompts.empty()) {
            return Status::OK();
        }

        std::string request_body;
        RETURN_IF_ERROR(adapter->build_embedding_request(batch_prompts, request_body));
        std::vector<std::vector<float>> batch_results;
        RETURN_IF_ERROR(_execute_prebuilt_embedding_request(
                request_body, batch_results, batch_prompts.size(), config, adapter, context));
        for (const auto& batch_result : batch_results) {
            _insert_embedding_result(col_result, batch_result);
        }
        batch_prompts.clear();
        return Status::OK();
    }

    // EMBED-private helper.
    // Flushes one accumulated multimodal embedding batch into the output array column.
    Status _flush_multimodal_embedding_batch(std::vector<MultimodalType>& batch_media_types,
                                             std::vector<std::string>& batch_media_content_types,
                                             std::vector<std::string>& batch_media_urls,
                                             ColumnArray& col_result, const AIResource& config,
                                             std::shared_ptr<AIAdapter>& adapter,
                                             FunctionContext* context) const {
        if (batch_media_urls.empty()) {
            return Status::OK();
        }

        std::string request_body;
        RETURN_IF_ERROR(adapter->build_multimodal_embedding_request(
                batch_media_types, batch_media_urls, batch_media_content_types, request_body));

        std::vector<std::vector<float>> batch_results;
        RETURN_IF_ERROR(_execute_prebuilt_embedding_request(
                request_body, batch_results, batch_media_urls.size(), config, adapter, context));
        for (const auto& batch_result : batch_results) {
            _insert_embedding_result(col_result, batch_result);
        }
        batch_media_types.clear();
        batch_media_content_types.clear();
        batch_media_urls.clear();
        return Status::OK();
    }

    static void _insert_embedding_result(ColumnArray& col_array,
                                         const std::vector<float>& float_result) {
        auto& offsets = col_array.get_offsets();
        auto& nested_nullable_col = assert_cast<ColumnNullable&>(col_array.get_data());
        auto& nested_col =
                assert_cast<ColumnFloat32&>(*(nested_nullable_col.get_nested_column_ptr()));
        nested_col.reserve(nested_col.size() + float_result.size());

        size_t current_offset = nested_col.size();
        nested_col.insert_many_raw_data(reinterpret_cast<const char*>(float_result.data()),
                                        float_result.size());
        offsets.push_back(current_offset + float_result.size());
        auto& null_map = nested_nullable_col.get_null_map_column();
        null_map.insert_many_vals(0, float_result.size());
    }

    static ColumnPtr _expand_and_wrap_nullable_result(ColumnArray::MutablePtr result,
                                                      ColumnUInt8::MutablePtr result_null_map,
                                                      size_t input_rows_count) {
        if (!result_null_map) {
            return result;
        }

        auto& offsets = result->get_offsets();
        size_t compact_row = offsets.size();
        offsets.resize(input_rows_count);
        // For example, embedding rows 1 and 3 produces compact offsets [5, 10]. Given
        // result_null_map [1, 0, 1, 0, 1], expand them to [0, 5, 5, 10, 10], where NULL rows
        // reuse the previous offset. Fill backwards to avoid overwriting unread compact offsets.
        for (size_t row = input_rows_count; row-- > 0;) {
            if (result_null_map->get_data()[row]) {
                offsets[row] = compact_row == 0 ? 0 : offsets[compact_row - 1];
            } else {
                offsets[row] = offsets[--compact_row];
            }
        }
        return ColumnNullable::create(std::move(result), std::move(result_null_map));
    }

    static bool _starts_with_ignore_case(std::string_view s, std::string_view prefix) {
        if (s.size() < prefix.size()) {
            return false;
        }
        return std::equal(prefix.begin(), prefix.end(), s.begin(), [](char a, char b) {
            return std::tolower(static_cast<unsigned char>(a)) ==
                   std::tolower(static_cast<unsigned char>(b));
        });
    }

    static Status _infer_media_type(const rapidjson::Value& file_input, std::string& content_type,
                                    MultimodalType& media_type) {
        RETURN_IF_ERROR(_get_required_string_field(file_input, "content_type", content_type));

        if (_starts_with_ignore_case(content_type, "image/")) {
            media_type = MultimodalType::IMAGE;
            return Status::OK();
        } else if (_starts_with_ignore_case(content_type, "video/")) {
            media_type = MultimodalType::VIDEO;
            return Status::OK();
        } else if (_starts_with_ignore_case(content_type, "audio/")) {
            media_type = MultimodalType::AUDIO;
            return Status::OK();
        }

        return Status::InvalidArgument("Unsupported content_type for EMBED: {}", content_type);
    }

    // Parse the FILE-like JSONB argument into a JSON object for downstream field reads.
    static Status _parse_file_input(const IColumn& file_column, size_t row_num,
                                    rapidjson::Document& file_input) {
        StringRef file_ref = file_column.get_data_at(row_num);
        std::string file_json = JsonbToJson::jsonb_to_json_string(file_ref.data, file_ref.size);
        file_input.Parse(file_json.c_str());
        DORIS_CHECK(!file_input.HasParseError() && file_input.IsObject());
        return Status::OK();
    }

    // TODO(lzq): After support FILE type, We should use the interface provided by FILE to get the fields
    // replacing this function
    static Status _get_required_string_field(const rapidjson::Value& obj, const char* field_name,
                                             std::string& value) {
        auto iter = obj.FindMember(field_name);
        if (iter == obj.MemberEnd() || !iter->value.IsString()) {
            return Status::InvalidArgument(
                    "EMBED file json field '{}' is required and must be a string", field_name);
        }
        value = iter->value.GetString();
        if (value.empty()) {
            return Status::InvalidArgument("EMBED file json field '{}' can not be empty",
                                           field_name);
        }
        return Status::OK();
    }

    static Status init_s3_client_conf_from_json(const rapidjson::Value& file_input,
                                                S3ClientConf& s3_client_conf) {
        std::string endpoint;
        RETURN_IF_ERROR(_get_required_string_field(file_input, "endpoint", endpoint));
        std::string region;
        RETURN_IF_ERROR(_get_required_string_field(file_input, "region", region));

        auto get_optional_string_field = [&](const char* field_name, std::string& value) {
            auto iter = file_input.FindMember(field_name);
            if (iter == file_input.MemberEnd() || iter->value.IsNull()) {
                return;
            }
            DORIS_CHECK(iter->value.IsString());
            value = iter->value.GetString();
        };

        get_optional_string_field("ak", s3_client_conf.ak);
        get_optional_string_field("sk", s3_client_conf.sk);
        get_optional_string_field("role_arn", s3_client_conf.role_arn);
        get_optional_string_field("external_id", s3_client_conf.external_id);
        s3_client_conf.endpoint = endpoint;
        s3_client_conf.region = region;

        return Status::OK();
    }

    Status _resolve_media_url(const rapidjson::Value& file_input, int64_t ttl_seconds,
                              std::string& media_url) const {
        std::string uri;
        RETURN_IF_ERROR(_get_required_string_field(file_input, "uri", uri));

        // If it's a direct http/https URL, use it as-is
        if (_starts_with_ignore_case(uri, "http://") || _starts_with_ignore_case(uri, "https://")) {
            media_url = uri;
            return Status::OK();
        }

        S3ClientConf s3_client_conf;
        RETURN_IF_ERROR(init_s3_client_conf_from_json(file_input, s3_client_conf));
        auto s3_client = DORIS_TRY(S3ClientFactory::instance().create(s3_client_conf));

        S3URI s3_uri(uri);
        RETURN_IF_ERROR(s3_uri.parse());
        std::string bucket = s3_uri.get_bucket();
        std::string key = s3_uri.get_key();
        DORIS_CHECK(!bucket.empty() && !key.empty());
        media_url = s3_client->generate_presigned_url({.bucket = bucket, .key = key, .prefix = ""},
                                                      ttl_seconds);
        return Status::OK();
    }
};

}; // namespace doris

Coverage Report

Created: 2026-09-22 20:00

Line	Count	Source
1		// Licensed to the Apache Software Foundation (ASF) under one
2		// or more contributor license agreements. See the NOTICE file
3		// distributed with this work for additional information
4		// regarding copyright ownership. The ASF licenses this file
5		// to you under the Apache License, Version 2.0 (the
6		// "License"); you may not use this file except in compliance
7		// with the License. You may obtain a copy of the License at
8		//
9		// http://www.apache.org/licenses/LICENSE-2.0
10		//
11		// Unless required by applicable law or agreed to in writing,
12		// software distributed under the License is distributed on an
13		// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14		// KIND, either express or implied. See the License for the
15		// specific language governing permissions and limitations
16		// under the License.
17
18		#pragma once
19
20		#include <glog/logging.h>
21		#include <rapidjson/document.h>
22
23		#include <string_view>
24
25		#include "core/data_type/data_type_nullable.h"
26		#include "core/data_type/primitive_type.h"
27		#include "exprs/function/ai/ai_functions.h"
28		#include "util/jsonb_utils.h"
29		#include "util/s3_uri.h"
30		#include "util/s3_util.h"
31
32		namespace doris {
33		class FunctionEmbed : public AIFunction<FunctionEmbed> {
34		public:
35		static constexpr auto name = "embed";
36
37		static constexpr size_t number_of_arguments = 2;
38
39		static constexpr auto system_prompt = "";
40
41	5	DataTypePtr get_nested_return_type_impl(const DataTypes& /arguments/) const {
42	5	return std::make_shared<DataTypeArray>(make_nullable(std::make_shared<DataTypeFloat32>()));
43	5	}
44
45		using PreparedFunctionImpl::execute;
46
47	23	AIResource select_ai_resource(const TAIResource& resource, PrimitiveType input_type) const {
48	23	bool has_complete_multimodal_embed_properties = _has_complete_resource_properties(
49	23	resource.embed_mm_endpoint, resource.embed_mm_provider_type,
50	23	resource.embed_mm_model_name, resource.embed_mm_api_key);
51	23	if (input_type == PrimitiveType::TYPE_JSONB && has_complete_multimodal_embed_properties) {
52	1	return AIResource::from_multimodal_embed(resource);
53	1	}
54	22	bool has_complete_embed_properties = _has_complete_resource_properties(
55	22	resource.embed_endpoint, resource.embed_provider_type, resource.embed_model_name,
56	22	resource.embed_api_key);
57	22	return has_complete_embed_properties ? AIResource::from_embed(resource)
58	22	: AIResource(resource);
59	23	}
60
61		Status execute(FunctionContext* context, Block& block, const ColumnNumbers& arguments,
62		uint32_t result, size_t input_rows_count, const AIResource& config,
63	18	std::shared_ptr<AIAdapter>& adapter) const {
64	18	if (arguments.size() != 2) {
65	1	return Status::InvalidArgument("Function EMBED expects 2 arguments, but got {}",
66	1	arguments.size());
67	1	}
68
69	17	const auto& input = block.get_by_position(arguments[1]);
70	17	ColumnUInt8::MutablePtr result_null_map;
71	17	if (input.type->is_nullable()) {
72	4	const auto& [column, is_const] = unpack_if_const(input.column);
73	4	const auto& nullable =
74	4	assert_cast<const ColumnNullable&, TypeCheckOnRelease::DISABLE>(*column);
75	4	result_null_map = ColumnUInt8::create(input_rows_count, 0);
76	4	VectorizedUtils::update_null_map(result_null_map->get_data(),
77	4	nullable.get_null_map_data(), is_const);
78	4	}
79
80	17	if (result_null_map &&
81	17	!simd::contain_zero(result_null_map->get_data().data(), input_rows_count)) {
82	1	block.get_by_position(result).column =
83	1	block.get_by_position(result).type->create_column_const(input_rows_count,
84	1	Field());
85	1	return Status::OK();
86	1	}
87
88	16	ColumnPtr input_column =
89	16	input.unnest_nullable(input.type->is_nullable() ? input.get_nullable_column_info()
90	16	: NullableColumnInfo {},
91	16	false)
92	16	.column;
93	16	PrimitiveType input_type = remove_nullable(input.type)->get_primitive_type();
94	16	if (input_type == PrimitiveType::TYPE_JSONB) {
95	9	return _execute_multimodal_embed(context, block, result, input_rows_count, config,
96	9	adapter, input_column, std::move(result_null_map));
97	9	}
98	7	if (input_type == PrimitiveType::TYPE_STRING \|\| input_type == PrimitiveType::TYPE_VARCHAR \|\|
99	7	input_type == PrimitiveType::TYPE_CHAR) {
100	6	return _execute_text_embed(context, block, result, input_rows_count, config, adapter,
101	6	input_column, std::move(result_null_map));
102	6	}
103	1	return Status::InvalidArgument(
104	1	"Function EMBED expects the second argument to be STRING or JSON, but got type {}",
105	1	input.type->get_name());
106	7	}
107
108	17	static FunctionPtr create() { return std::make_shared<FunctionEmbed>(); }
109
110		private:
111		static bool _has_complete_resource_properties(std::string_view endpoint,
112		std::string_view provider_type,
113		std::string_view model_name,
114	45	std::string_view api_key) {
115	45	return !endpoint.empty() && !provider_type.empty() && !model_name.empty() &&
116	45	(provider_type == "LOCAL" \|\| !api_key.empty());
117	45	}
118
119	15	static int32_t _get_embed_max_batch_size(FunctionContext* context) {
120	15	QueryContext* query_ctx = context->state()->get_query_ctx();
121	15	DORIS_CHECK(query_ctx != nullptr);
122
123	15	return query_ctx->query_options().embed_max_batch_size;
124	15	}
125
126		Status _execute_text_embed(FunctionContext* context, Block& block, uint32_t result,
127		size_t input_rows_count, const AIResource& config,
128		std::shared_ptr<AIAdapter>& adapter, const ColumnPtr& input_column,
129	6	ColumnUInt8::MutablePtr result_null_map) const {
130	6	auto col_result = ColumnArray::create(
131	6	ColumnNullable::create(ColumnFloat32::create(), ColumnUInt8::create()));
132	6	std::vector<std::string> batch_prompts;
133	6	size_t current_batch_size = 0;
134	6	const int32_t max_batch_size = _get_embed_max_batch_size(context);
135	6	const size_t max_context_window_size =
136	6	static_cast<size_t>(get_ai_context_window_size(context));
137	6	const NullMap* null_map = result_null_map ? &result_null_map->get_data() : nullptr;
138	6	const Columns prompt_columns {input_column};
139
140	29	for (size_t i = 0; i < input_rows_count; ++i) {
141	23	if (null_map && (*null_map)[i]) {
142	9	continue;
143	9	}
144
145	14	std::string prompt;
146	14	RETURN_IF_ERROR(build_prompt(prompt_columns, i, prompt));
147
148	14	const size_t prompt_size = prompt.size();
149
150	14	if (prompt_size > max_context_window_size) {
151		// flush history batch
152	0	RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
153	0	adapter, context));
154	0	current_batch_size = 0;
155
156	0	batch_prompts.emplace_back(std::move(prompt));
157	0	RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
158	0	adapter, context));
159	0	continue;
160	0	}
161
162	14	if (!batch_prompts.empty() &&
163	14	(current_batch_size + prompt_size > max_context_window_size \|\|
164	8	batch_prompts.size() >= static_cast<size_t>(max_batch_size))) {
165	3	RETURN_IF_ERROR(_flush_text_embedding_batch(batch_prompts, *col_result, config,
166	3	adapter, context));
167	3	current_batch_size = 0;
168	3	}
169
170	14	batch_prompts.emplace_back(std::move(prompt));
171	14	current_batch_size += prompt_size;
172	14	}
173
174	6	RETURN_IF_ERROR(
175	6	_flush_text_embedding_batch(batch_prompts, *col_result, config, adapter, context));
176
177	6	block.replace_by_position(result, _expand_and_wrap_nullable_result(
178	6	std::move(col_result), std::move(result_null_map),
179	6	input_rows_count));
180	6	return Status::OK();
181	6	}
182
183		Status _execute_multimodal_embed(FunctionContext* context, Block& block, uint32_t result,
184		size_t input_rows_count, const AIResource& config,
185		std::shared_ptr<AIAdapter>& adapter,
186		const ColumnPtr& input_column,
187	9	ColumnUInt8::MutablePtr result_null_map) const {
188	9	auto col_result = ColumnArray::create(
189	9	ColumnNullable::create(ColumnFloat32::create(), ColumnUInt8::create()));
190	9	std::vector<MultimodalType> batch_media_types;
191	9	std::vector<std::string> batch_media_content_types;
192	9	std::vector<std::string> batch_media_urls;
193	9	const NullMap* null_map = result_null_map ? &result_null_map->get_data() : nullptr;
194
195	9	int64_t ttl_seconds = 3600;
196	9	QueryContext* query_ctx = context->state()->get_query_ctx();
197	9	if (query_ctx && query_ctx->query_options().__isset.file_presigned_url_ttl_seconds) {
198	9	ttl_seconds = query_ctx->query_options().file_presigned_url_ttl_seconds;
199	9	if (ttl_seconds <= 0) {
200	1	ttl_seconds = 3600;
201	1	}
202	9	}
203
204	9	const int32_t max_batch_size = _get_embed_max_batch_size(context);
205
206	24	for (size_t i = 0; i < input_rows_count; ++i) {
207	19	if (null_map && (*null_map)[i]) {
208	3	continue;
209	3	}
210
211	16	rapidjson::Document file_input;
212	16	RETURN_IF_ERROR(_parse_file_input(*input_column, i, file_input));
213
214	16	std::string content_type;
215	16	MultimodalType media_type;
216	16	RETURN_IF_ERROR(_infer_media_type(file_input, content_type, media_type));
217
218	14	std::string media_url;
219	14	RETURN_IF_ERROR(_resolve_media_url(file_input, ttl_seconds, media_url));
220
221	12	if (!batch_media_urls.empty() &&
222	12	batch_media_urls.size() >= static_cast<size_t>(max_batch_size)) {
223	1	RETURN_IF_ERROR(_flush_multimodal_embedding_batch(
224	1	batch_media_types, batch_media_content_types, batch_media_urls, *col_result,
225	1	config, adapter, context));
226	1	}
227
228	12	batch_media_types.emplace_back(media_type);
229	12	batch_media_content_types.emplace_back(std::move(content_type));
230	12	batch_media_urls.emplace_back(std::move(media_url));
231	12	}
232
233	5	RETURN_IF_ERROR(_flush_multimodal_embedding_batch(
234	5	batch_media_types, batch_media_content_types, batch_media_urls, *col_result, config,
235	5	adapter, context));
236
237	5	block.replace_by_position(result, _expand_and_wrap_nullable_result(
238	5	std::move(col_result), std::move(result_null_map),
239	5	input_rows_count));
240	5	return Status::OK();
241	5	}
242
243		// EMBED-private helper.
244		// Sends one embedding request with a prebuilt request body and validates returned row count.
245		Status _execute_prebuilt_embedding_request(const std::string& request_body,
246		std::vector<std::vector<float>>& results,
247		size_t expected_size, const AIResource& config,
248		std::shared_ptr<AIAdapter>& adapter,
249	15	FunctionContext* context) const {
250	15	std::string response;
251	15	#ifdef BE_TEST
252	15	if (config.provider_type == "MOCK") {
253	15	results.clear();
254	15	results.reserve(expected_size);
255	41	for (size_t i = 0; i < expected_size; ++i) {
256	26	results.emplace_back(std::initializer_list<float> {0, 1, 2, 3, 4});
257	26	}
258	15	return Status::OK();
259	15	}
260	0	#endif
261
262	0	RETURN_IF_ERROR(
263	0	this->send_request_to_llm(request_body, response, config, adapter, context));
264
265	0	RETURN_IF_ERROR(adapter->parse_embedding_response(response, results));
266	0	if (results.empty()) {
267	0	return Status::InternalError("AI returned empty result");
268	0	}
269	0	if (results.size() != expected_size) [[unlikely]] {
270	0	return Status::InternalError(
271	0	"AI embedding returned {} results, but {} inputs were sent", results.size(),
272	0	expected_size);
273	0	}
274	0	return Status::OK();
275	0	}
276
277		// EMBED-private helper.
278		// Flushes one accumulated text embedding batch into the output array column.
279		Status _flush_text_embedding_batch(std::vector<std::string>& batch_prompts,
280		ColumnArray& col_result, const AIResource& config,
281		std::shared_ptr<AIAdapter>& adapter,
282	9	FunctionContext* context) const {
283	9	if (batch_prompts.empty()) {
284	0	return Status::OK();
285	0	}
286
287	9	std::string request_body;
288	9	RETURN_IF_ERROR(adapter->build_embedding_request(batch_prompts, request_body));
289	9	std::vector<std::vector<float>> batch_results;
290	9	RETURN_IF_ERROR(_execute_prebuilt_embedding_request(
291	9	request_body, batch_results, batch_prompts.size(), config, adapter, context));
292	14	for (const auto& batch_result : batch_results) {
293	14	_insert_embedding_result(col_result, batch_result);
294	14	}
295	9	batch_prompts.clear();
296	9	return Status::OK();
297	9	}
298
299		// EMBED-private helper.
300		// Flushes one accumulated multimodal embedding batch into the output array column.
301		Status _flush_multimodal_embedding_batch(std::vector<MultimodalType>& batch_media_types,
302		std::vector<std::string>& batch_media_content_types,
303		std::vector<std::string>& batch_media_urls,
304		ColumnArray& col_result, const AIResource& config,
305		std::shared_ptr<AIAdapter>& adapter,
306	6	FunctionContext* context) const {
307	6	if (batch_media_urls.empty()) {
308	0	return Status::OK();
309	0	}
310
311	6	std::string request_body;
312	6	RETURN_IF_ERROR(adapter->build_multimodal_embedding_request(
313	6	batch_media_types, batch_media_urls, batch_media_content_types, request_body));
314
315	6	std::vector<std::vector<float>> batch_results;
316	6	RETURN_IF_ERROR(_execute_prebuilt_embedding_request(
317	6	request_body, batch_results, batch_media_urls.size(), config, adapter, context));
318	12	for (const auto& batch_result : batch_results) {
319	12	_insert_embedding_result(col_result, batch_result);
320	12	}
321	6	batch_media_types.clear();
322	6	batch_media_content_types.clear();
323	6	batch_media_urls.clear();
324	6	return Status::OK();
325	6	}
326
327		static void _insert_embedding_result(ColumnArray& col_array,
328	26	const std::vector<float>& float_result) {
329	26	auto& offsets = col_array.get_offsets();
330	26	auto& nested_nullable_col = assert_cast<ColumnNullable&>(col_array.get_data());
331	26	auto& nested_col =
332	26	assert_cast<ColumnFloat32&>(*(nested_nullable_col.get_nested_column_ptr()));
333	26	nested_col.reserve(nested_col.size() + float_result.size());
334
335	26	size_t current_offset = nested_col.size();
336	26	nested_col.insert_many_raw_data(reinterpret_cast<const char*>(float_result.data()),
337	26	float_result.size());
338	26	offsets.push_back(current_offset + float_result.size());
339	26	auto& null_map = nested_nullable_col.get_null_map_column();
340	26	null_map.insert_many_vals(0, float_result.size());
341	26	}
342
343		static ColumnPtr _expand_and_wrap_nullable_result(ColumnArray::MutablePtr result,
344		ColumnUInt8::MutablePtr result_null_map,
345	11	size_t input_rows_count) {
346	11	if (!result_null_map) {
347	8	return result;
348	8	}
349
350	3	auto& offsets = result->get_offsets();
351	3	size_t compact_row = offsets.size();
352	3	offsets.resize(input_rows_count);
353		// For example, embedding rows 1 and 3 produces compact offsets [5, 10]. Given
354		// result_null_map [1, 0, 1, 0, 1], expand them to [0, 5, 5, 10, 10], where NULL rows
355		// reuse the previous offset. Fill backwards to avoid overwriting unread compact offsets.
356	24	for (size_t row = input_rows_count; row-- > 0;) {
357	21	if (result_null_map->get_data()[row]) {
358	12	offsets[row] = compact_row == 0 ? 0 : offsets[compact_row - 1];
359	12	} else {
360	9	offsets[row] = offsets[--compact_row];
361	9	}
362	21	}
363	3	return ColumnNullable::create(std::move(result), std::move(result_null_map));
364	11	}
365
366	52	static bool _starts_with_ignore_case(std::string_view s, std::string_view prefix) {
367	52	if (s.size() < prefix.size()) {
368	0	return false;
369	0	}
370	243	return std::equal(prefix.begin(), prefix.end(), s.begin(), [](char a, char b) {
371	243	return std::tolower(static_cast<unsigned char>(a)) ==
372	243	std::tolower(static_cast<unsigned char>(b));
373	243	});
374	52	}
375
376		static Status _infer_media_type(const rapidjson::Value& file_input, std::string& content_type,
377	16	MultimodalType& media_type) {
378	16	RETURN_IF_ERROR(_get_required_string_field(file_input, "content_type", content_type));
379
380	15	if (_starts_with_ignore_case(content_type, "image/")) {
381	9	media_type = MultimodalType::IMAGE;
382	9	return Status::OK();
383	9	} else if (_starts_with_ignore_case(content_type, "video/")) {
384	3	media_type = MultimodalType::VIDEO;
385	3	return Status::OK();
386	3	} else if (_starts_with_ignore_case(content_type, "audio/")) {
387	2	media_type = MultimodalType::AUDIO;
388	2	return Status::OK();
389	2	}
390
391	1	return Status::InvalidArgument("Unsupported content_type for EMBED: {}", content_type);
392	15	}
393
394		// Parse the FILE-like JSONB argument into a JSON object for downstream field reads.
395		static Status _parse_file_input(const IColumn& file_column, size_t row_num,
396	16	rapidjson::Document& file_input) {
397	16	StringRef file_ref = file_column.get_data_at(row_num);
398	16	std::string file_json = JsonbToJson::jsonb_to_json_string(file_ref.data, file_ref.size);
399	16	file_input.Parse(file_json.c_str());
400	16	DORIS_CHECK(!file_input.HasParseError() && file_input.IsObject());
401	16	return Status::OK();
402	16	}
403
404		// TODO(lzq): After support FILE type, We should use the interface provided by FILE to get the fields
405		// replacing this function
406		static Status _get_required_string_field(const rapidjson::Value& obj, const char* field_name,
407	35	std::string& value) {
408	35	auto iter = obj.FindMember(field_name);
409	35	if (iter == obj.MemberEnd() \|\| !iter->value.IsString()) {
410	3	return Status::InvalidArgument(
411	3	"EMBED file json field '{}' is required and must be a string", field_name);
412	3	}
413	32	value = iter->value.GetString();
414	32	if (value.empty()) {
415	0	return Status::InvalidArgument("EMBED file json field '{}' can not be empty",
416	0	field_name);
417	0	}
418	32	return Status::OK();
419	32	}
420
421		static Status init_s3_client_conf_from_json(const rapidjson::Value& file_input,
422	3	S3ClientConf& s3_client_conf) {
423	3	std::string endpoint;
424	3	RETURN_IF_ERROR(_get_required_string_field(file_input, "endpoint", endpoint));
425	2	std::string region;
426	2	RETURN_IF_ERROR(_get_required_string_field(file_input, "region", region));
427
428	4	auto get_optional_string_field = [&](const char* field_name, std::string& value) {
429	4	auto iter = file_input.FindMember(field_name);
430	4	if (iter == file_input.MemberEnd() \|\| iter->value.IsNull()) {
431	0	return;
432	0	}
433	4	DORIS_CHECK(iter->value.IsString());
434	4	value = iter->value.GetString();
435	4	};
436
437	1	get_optional_string_field("ak", s3_client_conf.ak);
438	1	get_optional_string_field("sk", s3_client_conf.sk);
439	1	get_optional_string_field("role_arn", s3_client_conf.role_arn);
440	1	get_optional_string_field("external_id", s3_client_conf.external_id);
441	1	s3_client_conf.endpoint = endpoint;
442	1	s3_client_conf.region = region;
443
444	1	return Status::OK();
445	2	}
446
447		Status _resolve_media_url(const rapidjson::Value& file_input, int64_t ttl_seconds,
448	14	std::string& media_url) const {
449	14	std::string uri;
450	14	RETURN_IF_ERROR(_get_required_string_field(file_input, "uri", uri));
451
452		// If it's a direct http/https URL, use it as-is
453	14	if (_starts_with_ignore_case(uri, "http://") \|\| _starts_with_ignore_case(uri, "https://")) {
454	11	media_url = uri;
455	11	return Status::OK();
456	11	}
457
458	3	S3ClientConf s3_client_conf;
459	3	RETURN_IF_ERROR(init_s3_client_conf_from_json(file_input, s3_client_conf));
460	1	auto s3_client = DORIS_TRY(S3ClientFactory::instance().create(s3_client_conf));
461
462	1	S3URI s3_uri(uri);
463	1	RETURN_IF_ERROR(s3_uri.parse());
464	1	std::string bucket = s3_uri.get_bucket();
465	1	std::string key = s3_uri.get_key();
466	1	DORIS_CHECK(!bucket.empty() && !key.empty());
467	1	media_url = s3_client->generate_presigned_url({.bucket = bucket, .key = key, .prefix = ""},
468	1	ttl_seconds);
469	1	return Status::OK();
470	1	}
471		};
472
473		}; // namespace doris