[ISSUE-255] Support natural language graph analytics by LLM

docs/docs-cn/application-development/dsl/chat/chat_guide.md

@@ -0,0 +1,25 @@
+# GQL大模型语法提示手册
+本手册列出了GQL常见的语法点以及参考的提示方法，用户可参考提问示例让模型生成对应的语法的GQL语句。
+
+| 语法 | 提问示例 | 结果 |
+| --- | --- | --- |
+| 查找点 | 请找出的类型为person**的点** | match(a:person) return a |
+| 查找边 | 返回图中所有的标签为knows**的边** | match(a)-[e:knows]->(b) return e |
+| 查找关系 | 查询10个属于北京的大学<br />找出5个和小张老师**有关系**的学生 | match(a:city where a.name = '北京')<-[:belong]-(b:university) return b limit 10<br />match(a:teacher where a.name='小张')-[e]-(b:student) return b limit 5 |
+| 查找多度关系 | 查找小王同学的朋友认识的人<br /><br />请帮我找出和标签为大学的点有拥有关系的部门,并且和这些部门有拥有关系的学生,并且和这些学生有选择关系的课程<br /><br />请查找腾讯和谷歌公司共同创建的软件的名称，返回5个。 | match(a:student where a.name = '小王')-[e:friend]->(b)-[e2:knows]->(c:person) return c<br /><br />match(a:university)-[e:has]->(b:department)-[e2:has]->(c:student)-[e3:selects]->(d:course) return d<br /><br />match(a:company where a.name='腾讯')-[e:creates]->(b:software)<-[e2:creates]-(c:company where c.name='谷歌') return b.name limit 5 |
+| Loop | 查询从person张思琪出发，经过类型为支付的边，**走2到4度到达的点** | match(a:person where a.name='张思琪')-[e:pay]->{2,4}(b:person) return b |
+| Loop(环) | 找出与person李红有knows关系的**3跳以内的环** | match(a:person where name = '李红')-[e:knows]->{1,2}(b)->(a) return a.id, b.id as b_id<br /> |
+| 过滤条件 | 找出小红认识的年龄大于20，工资大于5000的人<br /><br /><br />请帮我找出10个性别不是famale且身高小于160,或id大于5的节点 | match(a:person where a.name='小红')-[e:knows]->(b:person where b.age > 20 and b.salary > 5000) return b<br /><br />match(a where (a.gender <> 'female' and a.height < 160) or a.id > 5) return a limit 10 |
+| Let单值 | 查询蚂蚁集团创造的软件, **令**软件的minPrice**等于**软件的价格的最小值, 返回公司的id, 软件的minPrice | match(a:company where a.name = '蚂蚁集团')-[e:creates]->(b:software) let b.minPrice = MIN(b.price) return a.id, b.minPrice |
+| Let子查询 | 请帮我查找公司蚂蚁集团雇佣的person, **令**这个person的countSalary**的值等于**认识他的人的薪资的和, 再查找他购买的软件<br /><br />给出id为10的城市所属于的国家, 并将国家有关的公司的人数的平均值**赋值**给国家的avgCnt | match(a:company where a.name = '蚂蚁集团')-[e:employee]->(b:person) let b.countSalary = SUM((b:person)<-[e2:knows]-(c:person) => c.salary) match(b:person)-[e3:buy]->(d:software) return b.countSalary, d<br /><br />match(a:city where id = '10')-[e:belong]->(b:country)<-[e2:belong]-(c:company) let b.avgCnt = AVG(c.peopleNumber) return b |
+| 函数调用 | **调用SSSP函数**,以'arg1', 10**作为输入**,返回id, distance | match(a:person) call sssp(a, 10) yield (id, distance) return id, distance |
+| order | 返回公司创造的软件，并根据公司的规格从大到小、软件的价格**从小到大排序** | match(a:company)-[e:creates]->(b:software) return a.scale,b.price order by a.scale desc, b.price asc |
+| group by  | 找出小红认识的人，**根据**这些人的性别**分组**，返回工资的最大值<br /><br />帮我找出与北京大学有关联的公司，返回公司**以**规模**进行分组**的人数的平均值 | match(a:person where person.name = '小红')-[e:knows]->(b:person) return MAX(b.salary) group by b.gender<br /><br />match(a:university where a.name='北京大学')-[e]-(b:company) return AVG(b.peopleNumber) group by b.scale |
+| join | 找出郑伟喜欢的所有人，以及认识郑伟的所有人，并将它们**一起返回**<br /><br />请帮我找出和person alice有关联的学校,**称为X**,再帮我找出**和这个X有关联**的其他公司,以及和X有关联的person | match(a:person where a.name = '郑伟')-[e:likes]->(b:person),(a:person where a.name = '郑伟')<-[e2:knows]-(c:person) return a, b, c<br /><br />match(a:person where a.name = 'alice')-[e]-(b:school), (b:school)-[e2]-(c:company),(b:school)-[e3]-(d:person) return a, b, c, d |
+| 带图的schema单条查询<br />**（在Console中开启“附带图schema”开关会自动拼接，不需要用户自己输入**） | 使用这个图:CREATE GRAPH g ( Vertex film ( id int ID, name varchar, category varchar, value int ), Vertex cinema ( id int ID, name varchar, address varchar, size int ), Vertex person ( id int ID, name varchar, age int, gender varchar, height int, salary int ), Vertex comment ( id int ID, name varchar, createTime bigint, wordCount int ), Vertex tag ( id int ID, name varchar, value int ), Edge person_likes_comment ( srcId int FROM person SOURCE ID, targetId int FROM comment DESTINATION ID, weight double, f0 int, f1 boolean ), Edge cinema_releases_film ( srcId int FROM cinema SOURCE ID, targetId int FROM film DESTINATION ID, weight double, f0 int, f1 boolean ), Edge person_watch_film ( srcId int FROM person SOURCE ID, targetId int FROM film DESTINATION ID, weight double, f0 int, f1 boolean, timeStamp bigint ), Edge film_has_tag ( srcId int FROM film SOURCE ID, targetId int FROM tag DESTINATION ID, weight double, f0 int, f1 boolean ), Edge person_creates_comment ( srcId int FROM person SOURCE ID, targetId int FROM comment DESTINATION ID, weight double, f0 int, f1 boolean, timeStamp bigint ), Edge comment_belong_film ( srcId int FROM comment SOURCE ID, targetId int FROM film DESTINATION ID, weight double, f0 int, f1 boolean ));找出和person孙梅有person_creates_comment关系的comment,以及和person孙建聪有person_likes_comment关系的comment,将它们都返回 | atch(a:person where a.name = '孙梅')-[e:person_creates_comment]->(b:comment),(c:person where c.name = '孙建聪')-[e2:person_likes_comment]->(d:comment)return a, b, c, d |
+| 带图的schema多条查询 | 使用这个图:CREATE GRAPH g ( Vertex book ( id int ID, name varchar, id int ID, name varchar, category varchar, price int, wordCount int, createTime bigint ), Vertex publisher ( id int ID, name varchar, age int, gender varchar, height int, salary int ), Vertex reader ( id int ID, name varchar, age int, gender varchar, height int, salary int ), Vertex author ( id int ID, name varchar, age int, gender varchar, height int, salary int ), Edge author_write_book ( srcId int FROM author SOURCE ID, targetId int FROM book DESTINATION ID, weight double, f0 int, f1 boolean, timeStamp bigint ), Edge publisher_publish_book ( srcId int FROM publisher SOURCE ID, targetId int FROM book DESTINATION ID, weight double, f0 int, f1 boolean, timeStamp bigint ), Edge book_refers_book ( srcId int FROM book SOURCE ID, targetId int FROM book DESTINATION ID, weight double, f0 int, f1 boolean ), Edge reader_likes_book ( srcId int FROM reader SOURCE ID, targetId int FROM book DESTINATION ID, weight double, f0 int, f1 boolean ), Edge author_knows_author ( srcId int FROM author SOURCE ID, targetId int FROM author DESTINATION ID, weight double, f0 int, f1 boolean ));执行以下4个查询:1: 找出被作家黄建聪认识的作家;2: 返回源点标签为作家,目标点标签为作家,标签为author_knows_author的边;3: 请帮我查找所有与book计算机网络有关系的book的id;4: 请帮我查找152个与何雪和张建聪都存在关系的书节点; | 查询语句为:1: match(a:author)<-[e:author_knows_author]-(b:author where b.name='黄建聪') return a, b;2: match(a:author)-[e:author_knows_author]->(b:author) return e;3: match(a:book where a.name='计算机网络')-[e]-(b:book) return b.id;4: match(a where a.name='何雪')-[e]->(b:book)<-[e2]-(c where c.name='张建聪') return b limit 152; |
+
+
+
+
+

docs/docs-cn/deploy/install_llm.md

@@ -0,0 +1,80 @@
+用户可以在本地完成大模型的服务化部署，以下步骤描述了从下载预训练模型，服务化部署和调试的整个过程。用户机器确保已安装Docker，可访问大模型存储库。
+
+ ## 步骤1:下载大模型文件
+
+我们已将预训练好的大模型文件，上传至[Hugging Face仓库](https://huggingface.co/tugraph/CodeLlama-7b-GQL-hf)，下载模型文件后解压到本地。下载完成后，解压模型文件到指定的本地目录，如 /home/huggingface。
+![hugging](../../static/img/llm_hugging_face.png) 
+
+ ## 步骤2:准备 Docker 容器环境
+1. 在终端运行以下命令下载模型服务化所需的 Docker 镜像：
+
+```
+docker pull tugraph/llam_infer_service:0.0.1
+
+//使用以下命令来验证镜像是否成功下载
+docker images
+```
+
+2. 运行以下命令启动Docker容器：
+
+```
+docker run -it  --name ${容器名称} -v ${本地模型路径}:${容器模型路径} -p ${本地端口}:${容器服务化端口} -d ${镜像名}  
+
+//例如
+docker run -it --name my-model-container -v /home/huggingface:/opt/huggingface -p 8000:8000 -d llama_inference_server:v1
+
+// 检查容器是否正常运行
+docker ps 
+```
+
+在这里，我们将容器的 8000 端口映射到本地机器的 8000 端口，将本地模型（/home/huggingface）所在目录挂载到容器的路径（/opt/huggingface），并将容器名称设置为 my-model-container。
+
+## 步骤3:模型服务化部署
+1. 模型转换
+```
+//进入上述刚创建好的容器
+docker exec -it ${container_id} bash
+
+//执行如下命令
+cd /opt/llama_cpp
+python3 ./convert.py ${容器模型路径}
+```
+执行完成后， 会在容器模型路径下生成前缀为ggml-model的文件
+![undefined](../../static/img/llm_ggml_model.png) 
+
+2. 模型量化（可选）
+以llam2-7B模型为例：通过convert.py默认转换后的模型精度为F16，模型大小为13.0GB。如果当前机器资源无法满足这么大的模型推理，可通过./quantize对转换后的模型进一步量化。
+```
+// 如下所示 q4_0即将原始模型量化为int4，模型大小压缩至3.5GB；
+// 其它参数类似
+cd /opt/llama_cpp
+./quantize 默认生成的F16模型路径 量化后模型路径 q4_0
+```
+以下是量化后模型的大小和推理速度等参考指标：
+![undefined](../../static/img/llm_quantization_table.png)
+
+3. 模型服务化
+通过以下命令，将上述生成的模型进行服务化的部署，通过参数指定服务绑定的地址和端口：
+```
+// ./server -h 可查看参数详情
+// ${ggml-model...文件}为生成的ggml-model前缀的文件名
+
+cd /opt/llama_cpp
+./server --host ${ip} --port ${port} -m ${容器模型路径}/${ggml-model...文件} -c 4096
+
+//例如
+./server --host 0.0.0.0 --port 8000 -m  /opt/huggingface/ggml-model-f16.gguf -c 4096
+```
+
+4. 调试服务
+向服务地址发送http请求，其中"prompt"为查询的语句, 在返回的"content"中则是推理结果。
+
+```
+curl --request POST \
+    --url http://127.0.0.1:8000/completion \
+    --header "Content-Type: application/json" \
+    --data '{"prompt": "请返回小红的10个年龄大于20的朋友","n_predict": 128}'
+```
+如下是服务化部署后的模型推理结果：
+![undefined](../../static/img/llm_chat_result.png)
+

docs/docs-en/deploy/install_llm.md

@@ -0,0 +1,83 @@
+The users have the capability to locally deploy extensive models as a service. The complete process, encompassing downloading pre-trained models, deploying them as a service, and debugging, is described in the following steps. It is essential for the user's machine to have Docker installed and be granted access to the repository containing these large models.
+
+ ## Step 1: Download the model file
+ The pre-trained large model file has been uploaded to the [Hugging Face repository](https://huggingface.co/tugraph/CodeLlama-7b-GQL-hf). Please proceed with downloading and locally unzipping the model file.
+![hugging](../../static/img/llm_hugging_face.png)
+
+ ## Step 2: Prepare the Docker container environment
+1. Run the following command on the terminal to download the Docker image required for model servicing:
+
+```
+docker pull tugraph/llam_infer_service:0.0.1
+
+// Use the following command to verify that the image was successfully downloaded
+
+docker images
+```
+
+2. Run the following command to start the Docker container:
+
+```
+docker run -it  --name ${Container name} -v ${Local model path}:${Container model path} -p ${Local port}:${Container service port} -d ${Image name}  
+
+// Such as
+docker run -it --name my-model-container -v /home/huggingface:/opt/huggingface -p 8000:8000 -d llama_inference_server:v1
+
+// Check whether the container is running properly
+docker ps 
+```
+
+Here, we map the container's port 8000 to the local machine's port 8000, mount the directory where the local model (/home/huggingface) resides to the container's path (/opt/huggingface), and set the container name to my-model-container.
+
+## Step 3: Model service deployment
+1. Model transformation
+```
+// Enter the container you just created
+docker exec -it ${container_id} bash
+
+// Execute the following command
+cd /opt/llama_cpp
+python3 ./convert.py ${Container model path}
+```
+When the execution is complete, a file with the prefix ggml-model is generated under the container model path.
+![undefined](../../static/img/llm_ggml_model.png)
+
+2. Model quantization (optional)
+Take the llam2-7B model as an example: By default, the accuracy of the model converted by convert.py is F16 and the model size is 13.0GB. If the current machine resources cannot satisfy such a large model inference, the converted model can be further quantized by./quantize.
+
+```
+// As shown below, q4_0 quantizes the original model to int4 and compresses the model size to 3.5GB
+
+cd /opt/llama_cpp
+./quantize ${Default generated F16 model path} ${Quantized model path} q4_0
+```
+The following are reference indicators such as the size and reasoning speed of the quantized model:
+![undefined](../../static/img/llm_quantization_table.png)
+
+3. Model servicing
+Run the following command to deploy the above generated model as a service, and specify the address and port of the service binding through the parameters:
+```
+// ./server -h. You can view parameter details
+// ${ggml-model...file} The file name prefixes the generated ggml-model
+
+cd /opt/llama_cpp
+./server --host ${ip} --port ${port} -m ${Container model path}/${ggml-model...file} -c 4096
+
+// Such as
+./server --host 0.0.0.0 --port 8000 -m  /opt/huggingface/ggml-model-f16.gguf -c 4096
+```
+
+4. Debugging service
+Send an http request to the service address, where "prompt" is the query statement and "content" is the inference result.
+
+```
+curl --request POST \
+    --url http://127.0.0.1:8000/completion \
+    --header "Content-Type: application/json" \
+    --data '{"prompt": "请返回小红的10个年龄大于20的朋友","n_predict": 128}'
+```
+Debugging service
+The following is the model inference result after service deployment:
+![undefined](../../static/img/llm_chat_result.png)
+
+

geaflow-console/app/biz/shared/src/main/java/com/antgroup/geaflow/console/biz/shared/ChatManager.java

@@ -0,0 +1,28 @@
+/*
+ * Copyright 2023 AntGroup CO., Ltd.
+ *
+ * Licensed 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.
+ */
+
+package com.antgroup.geaflow.console.biz.shared;
+
+
+import com.antgroup.geaflow.console.biz.shared.view.ChatView;
+import com.antgroup.geaflow.console.common.dal.model.ChatSearch;
+
+public interface ChatManager extends IdManager<ChatView, ChatSearch> {
+
+    String callSync(ChatView chatView, boolean record, boolean withSchema);
+
+    String callASync(ChatView chatView, boolean withSchema);
+
+    boolean dropByJobId(String jobId);
+}

geaflow-console/app/biz/shared/src/main/java/com/antgroup/geaflow/console/biz/shared/LLMManager.java

@@ -0,0 +1,26 @@
+/*
+ * Copyright 2023 AntGroup CO., Ltd.
+ *
+ * Licensed 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.
+ */
+
+package com.antgroup.geaflow.console.biz.shared;
+
+
+import com.antgroup.geaflow.console.biz.shared.view.LLMView;
+import com.antgroup.geaflow.console.common.dal.model.LLMSearch;
+import com.antgroup.geaflow.console.common.util.type.GeaflowLLMType;
+import java.util.List;
+
+public interface LLMManager extends NameManager<LLMView, LLMSearch> {
+
+    List<GeaflowLLMType> getLLMTypes();
+}

geaflow-console/app/biz/shared/src/main/java/com/antgroup/geaflow/console/biz/shared/convert/ChatViewConverter.java

@@ -0,0 +1,55 @@
+/*
+ * Copyright 2023 AntGroup CO., Ltd.
+ *
+ * Licensed 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.
+ */
+
+package com.antgroup.geaflow.console.biz.shared.convert;
+
+import com.antgroup.geaflow.console.biz.shared.view.ChatView;
+import com.antgroup.geaflow.console.core.model.llm.GeaflowChat;
+import org.springframework.stereotype.Component;
+
+@Component
+public class ChatViewConverter extends IdViewConverter<GeaflowChat, ChatView> {
+
+    @Override
+    public void merge(ChatView view, ChatView updateView) {
+        super.merge(view, updateView);
+    }
+
+    @Override
+    protected ChatView modelToView(GeaflowChat model) {
+        ChatView view = super.modelToView(model);
+        view.setModelId(model.getModelId());
+        view.setAnswer(model.getAnswer());
+        view.setPrompt(model.getPrompt());
+        view.setStatus(model.getStatus());
+        view.setJobId(model.getJobId());
+        return view;
+    }
+
+    @Override
+    protected GeaflowChat viewToModel(ChatView view) {
+        GeaflowChat model = super.viewToModel(view);
+        model.setModelId(view.getModelId());
+        model.setAnswer(view.getAnswer());
+        model.setPrompt(view.getPrompt());
+        model.setStatus(view.getStatus());
+        model.setJobId(view.getJobId());
+        return model;
+    }
+
+    public GeaflowChat convert(ChatView view) {
+        return viewToModel(view);
+    }
+
+}