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) const {
        bool has_embed_properties = resource.__isset.embed_endpoint ||
                resource.__isset.embed_provider_type || resource.__isset.embed_model_name ||
                resource.__isset.embed_api_key;
        return has_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 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: 2026-09-10 01:11

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