feat: support multi-conversation memory #28

2026-02-12 18:29:21 +08:00 · 2023-02-05 18:38:57 +08:00
parent b0e81aeb0d
commit 7425d9075a
3 changed files with 88 additions and 8 deletions
--- a/README.md
+++ b/README.md
@@ -9,9 +9,11 @@
 - [x] **规则定制化：** 支持私聊中按指定规则触发自动回复，支持对群组设置自动回复白名单
 - [x] **多账号：** 支持多微信账号同时运行
 - [x] **图片生成：** 支持根据描述生成图片，并自动发送至个人聊天或群聊
+- [x] **上下文记忆**：支持多轮对话记忆，且为每个好友维护独立的上下会话


 # 更新日志
+>**2022.02.05：** 在openai官方接口方案中 (GPT-3模型) 实现上下文对话

 >**2022.12.19：** 引入 [itchat-uos](https://github.com/why2lyj/ItChat-UOS) 替换 itchat，解决由于不能登录网页微信而无法使用的问题，且解决Python3.9的兼容问题

@@ -85,7 +87,8 @@ cp config-template.json config.json
  "single_chat_reply_prefix": "[bot] ",                       # 私聊时自动回复的前缀，用于区分真人
  "group_chat_prefix": ["@bot"],                              # 群聊时包含该前缀则会触发机器人回复
  "group_name_white_list": ["ChatGPT测试群", "ChatGPT测试群2"], # 开启自动回复的群名称列表
-  "image_create_prefix": ["画", "看", "找"]                    # 开启图片回复的前缀
+  "image_create_prefix": ["画", "看", "找"],                   # 开启图片回复的前缀
+  "conversation_max_tokens": 3000                                                               # 支持上下文记忆的最多字符数
 }
 ```
 **配置说明：**
@@ -105,6 +108,7 @@ cp config-template.json config.json

 + 对于图像生成，在满足个人或群组触发条件外，还需要额外的关键词前缀来触发，对应配置 `image_create_prefix `
 + 关于OpenAI对话及图片接口的参数配置（内容自由度、回复字数限制、图片大小等），可以参考 [对话接口](https://beta.openai.com/docs/api-reference/completions) 和 [图像接口](https://beta.openai.com/docs/api-reference/completions)  文档直接在 [代码](https://github.com/zhayujie/chatgpt-on-wechat/blob/master/bot/openai/open_ai_bot.py) `bot/openai/open_ai_bot.py` 中进行调整。
+ `conversation_max_tokens`：表示能够记忆的上下文最大字数（一问一答为一组对话，如果累积的对话字数超出限制，就会优先移除最早的一组对话）


 ## 运行
--- a/bot/openai/open_ai_bot.py
+++ b/bot/openai/open_ai_bot.py
@@ -5,6 +5,7 @@ from config import conf
 from common.log import logger
 import openai

+user_session = dict()

 # OpenAI对话模型API (可用)
 class OpenAIBot(Bot):
@@ -12,17 +13,26 @@ class OpenAIBot(Bot):
        openai.api_key = conf().get('open_ai_api_key')

    def reply(self, query, context=None):
-        # auto append question mark
-        query = self.append_question_mark(query)

        # acquire reply content
        if not context or not context.get('type') or context.get('type') == 'TEXT':
-            return self.reply_text(query)
+            logger.info("[OPEN_AI] query={}".format(query))
+            from_user_id = context['from_user_id']
+            if query == '#清除记忆':
+                Session.clear_session(from_user_id)
+                return '记忆已清除'
+
+            new_query = Session.build_session_query(query, from_user_id)
+            logger.debug("[OPEN_AI] session query={}".format(new_query))
+
+            reply_content = self.reply_text(new_query, query)
+            Session.save_session(query, reply_content, from_user_id)
+            return reply_content
+
        elif context.get('type', None) == 'IMAGE_CREATE':
            return self.create_img(query)

-    def reply_text(self, query):
-        logger.info("[OPEN_AI] query={}".format(query))
+    def reply_text(self, query, origin_query):
        try:
            response = openai.Completion.create(
                model="text-davinci-003",  # 对话模型的名称
@@ -34,7 +44,7 @@ class OpenAIBot(Bot):
                presence_penalty=0.0,  # [-2,2]之间，该值越大则更倾向于产生不同的内容
                stop=["#"]
            )
-            res_content = response.choices[0]["text"].strip()
+            res_content = response.choices[0]["text"].strip().rstrip("<|im_end|>")
        except Exception as e:
            logger.exception(e)
            return None
@@ -93,3 +103,68 @@ class OpenAIBot(Bot):
            if query.endswith(symbol):
                return query
        return query + "?"
+
+
+class Session(object):
+    @staticmethod
+    def build_session_query(query, user_id):
+        '''
+        build query with conversation history
+        e.g.  Q: xxx
+              A: xxx
+              Q: xxx
+        :param query: query content
+        :param user_id: from user id
+        :return: query content with conversaction
+        '''
+        new_query = ""
+        session = user_session.get(user_id, None)
+        if session:
+            for conversation in session:
+                new_query += "Q: " + conversation["question"] + "\n\n\nA: " + conversation["answer"] + "<|im_end|>\n"
+            new_query += "Q: " + query + "\nA: "
+            return new_query
+        else:
+            return "Q: " + query + "\nA: "
+
+    @staticmethod
+    def save_session(query, answer, user_id):
+        max_tokens = conf().get("conversation_max_tokens")
+        if not max_tokens:
+            # default 3000
+            max_tokens = 3000
+        conversation = dict()
+        conversation["question"] = query
+        conversation["answer"] = answer
+        session = user_session.get(user_id)
+        if session:
+            # append conversation
+            session.append(conversation)
+        else:
+            # create session
+            queue = list()
+            queue.append(conversation)
+            user_session[user_id] = queue
+
+        # discard exceed limit conversation
+        Session.discard_exceed_conversation(user_session[user_id], max_tokens)
+
+
+    @staticmethod
+    def discard_exceed_conversation(session, max_tokens):
+        count = 0
+        count_list = list()
+        for i in range(len(session)-1, -1, -1):
+            # count tokens of conversation list
+            history_conv = session[i]
+            count += len(history_conv["question"]) + len(history_conv["answer"])
+            count_list.append(count)
+
+        for c in count_list:
+            if c > max_tokens:
+                # pop first conversation
+                session.pop(0)
+
+    @staticmethod
+    def clear_session(user_id):
+        user_session[user_id] = []
--- a/config-template.json
+++ b/config-template.json
@@ -4,5 +4,6 @@
  "single_chat_reply_prefix": "[bot] ",
  "group_chat_prefix": ["@bot"],
  "group_name_white_list": ["ChatGPT测试群", "ChatGPT测试群2"],
-  "image_create_prefix": ["画", "看", "找"]
+  "image_create_prefix": ["画", "看", "找"],
+  "conversation_max_tokens": 3000
 }