Experience Selection [Ready] (#406)

1. add dynamic experience selection by setting signal attachments. --------- Co-authored-by: Jack-Q <[email protected]>
microsoft · Sep 14, 2024 · 8885c90 · 8885c90
1 parent d6ae087
commit 8885c90
Show file tree

Hide file tree

Showing 39 changed files with 2,375 additions and 356 deletions.
diff --git a/README.md b/README.md
@@ -23,14 +23,16 @@ Unlike many agent frameworks that only track the chat history with LLMs in text,
 
 
 ## 🆕 News
+- 📅2024-09-13: We introduce the shared memory to store information that is shared between the roles in TaskWeaver. Please check the [memory](https://microsoft.github.io/TaskWeaver/docs/memory) for more details.🧠
+- 📅2024-09-13: We have enhanced the experience feature by allowing static and dynamic experience selection. Please check the [experience](https://microsoft.github.io/TaskWeaver/blog/experience) for more details.📚 
 - 📅2024-07-02: We have optimized TaskWeaver to support not-that-large language models served locally. Please check this [post](https://microsoft.github.io/TaskWeaver/blog/local_llm) for more details.🔗
 - 📅2024-05-07: We have added two blog posts on [Evaluating a LLM agent](https://microsoft.github.io/TaskWeaver/blog/evaluation) and [Adding new roles to TaskWeaver](https://microsoft.github.io/TaskWeaver/blog/role) in the documentation.📝
 - 📅2024-03-28: TaskWeaver now offers all-in-one Docker image, providing a convenient one-stop experience for users. Please check the [docker](https://microsoft.github.io/TaskWeaver/docs/usage/docker) for more details.🐳
 - 📅2024-03-27: TaskWeaver now switches to `container` mode by default for code execution. Please check the [code execution](https://microsoft.github.io/TaskWeaver/docs/code_execution) for more details.🐳
 - 📅2024-03-07: TaskWeaver now supports configuration of different LLMs for various components, such as the Planner and CodeInterpreter. Please check the [multi-llm](https://microsoft.github.io/TaskWeaver/docs/llms/multi-llm) for more details.🔗
 - 📅2024-03-04: TaskWeaver now supports a [container](https://microsoft.github.io/TaskWeaver/docs/code_execution) mode, which provides a more secure environment for code execution.🐳
 - 📅2024-02-28: TaskWeaver now offers a [CLI-only](https://microsoft.github.io/TaskWeaver/docs/advanced/cli_only) mode, enabling users to interact seamlessly with the Command Line Interface (CLI) using natural language.📟
-- 📅2024-02-01: TaskWeaver now has a plugin [document_retriever](https://github.com/microsoft/TaskWeaver/blob/main/project/plugins/README.md#document_retriever) for RAG based on a knowledge base.📚
+<!-- - 📅2024-02-01: TaskWeaver now has a plugin [document_retriever](https://github.com/microsoft/TaskWeaver/blob/main/project/plugins/README.md#document_retriever) for RAG based on a knowledge base.📚 -->
 <!-- - 📅2024-01-30: TaskWeaver introduces a new plugin-only mode that securely generates calls to specified plugins without producing extraneous code.🪡 --> 
 <!-- - 📅2024-01-23: TaskWeaver can now be personalized by transforming your chat histories into enduring [experiences](https://microsoft.github.io/TaskWeaver/docs/customization/experience) 🎉 -->
 <!-- - 📅2024-01-17: TaskWeaver now has a plugin [vision_web_explorer](https://github.com/microsoft/TaskWeaver/blob/main/project/plugins/README.md#vision_web_explorer) that can open a web browser and explore websites.🌐 -->

diff --git a/scripts/experience_mgt.py b/scripts/experience_mgt.py
diff --git a/taskweaver/code_interpreter/code_interpreter/code_generator.py b/taskweaver/code_interpreter/code_interpreter/code_generator.py
@@ -1,7 +1,7 @@
 import datetime
 import json
 import os
-from typing import List, Optional
+from typing import List, Optional, Tuple
 
 from injector import inject
 
@@ -55,8 +55,6 @@ def _configure(self) -> None:
         )
         self.auto_plugin_selection_topk = self._get_int("auto_plugin_selection_topk", 3)
 
-        self.use_experience = self._get_bool("use_experience", False)
-
         self.llm_alias = self._get_str("llm_alias", default="", required=False)
 
 
@@ -104,14 +102,7 @@ def __init__(
             logger.info("Plugin embeddings loaded")
             self.selected_plugin_pool = SelectedPluginPool()
 
-        if self.config.use_experience:
-            self.experience_generator = experience_generator
-            self.experience_generator.refresh()
-            self.experience_generator.load_experience()
-            self.logger.info(
-                "Experience loaded successfully, "
-                "there are {} experiences".format(len(self.experience_generator.experience_list)),
-            )
+        self.experience_generator = experience_generator
 
         self.logger.info("CodeGenerator initialized successfully")
 
@@ -166,15 +157,12 @@ def compose_prompt(
         self,
         rounds: List[Round],
         plugins: List[PluginEntry],
-        selected_experiences: Optional[List[Experience]] = None,
+        selected_experiences: Optional[List[Tuple[Experience, float]]] = None,
+        planning_enrichments: Optional[List[str]] = None,
     ) -> List[ChatMessageType]:
-        experiences = (
-            self.experience_generator.format_experience_in_prompt(
-                self.prompt_data["experience_instruction"],
-                selected_experiences,
-            )
-            if self.config.use_experience
-            else ""
+        experiences = self.format_experience(
+            template=self.prompt_data["experience_instruction"],
+            experiences=selected_experiences,
         )
 
         chat_history = [
@@ -207,6 +195,7 @@ def compose_prompt(
                 add_requirements=True,
                 summary=summary,
                 plugins=plugins,
+                planning_enrichments=planning_enrichments,
             ),
         )
         return chat_history
@@ -223,6 +212,7 @@ def compose_conversation(
         plugins: List[PluginEntry],
         add_requirements: bool = False,
         summary: Optional[str] = None,
+        planning_enrichments: Optional[List[str]] = None,
     ) -> List[ChatMessageType]:
         chat_history: List[ChatMessageType] = []
         ignored_types = [
@@ -260,18 +250,16 @@ def compose_conversation(
                         user_query = conversation_round.user_query
                         enrichment = f"The user request is: {user_query}\n\n"
 
-                        supplementary_info_dict = conversation_round.read_board()
-                        supplementary_info = "\n\n".join([bulletin for bulletin in supplementary_info_dict.values()])
-                        if supplementary_info != "":
-                            enrichment += f"Additional context:\n" f" {supplementary_info}\n\n"
+                        if planning_enrichments:
+                            enrichment += "Additional context:\n" + "\n".join(planning_enrichments) + "\n\n"
 
                     user_feedback = "None"
                     if last_post is not None and last_post.send_from == self.alias:
                         user_feedback = format_code_feedback(last_post)
 
                     user_message += self.user_message_head_template.format(
                         FEEDBACK=user_feedback,
-                        MESSAGE=f"{enrichment}{post.message}",
+                        MESSAGE=f"{enrichment}The task for this specific step is: {post.message}",
                     )
                 elif post.send_from == post.send_to == self.alias:
                     # for code correction
@@ -370,12 +358,24 @@ def reply(
         if self.config.enable_auto_plugin_selection:
             self.plugin_pool = self.select_plugins_for_prompt(query)
 
-        if self.config.use_experience:
-            selected_experiences = self.experience_generator.retrieve_experience(query)
+        exp_sub_paths = memory.get_shared_memory_entries(entry_type="experience_sub_path")
+
+        if exp_sub_paths:
+            self.tracing.set_span_attribute("experience_sub_path", str(exp_sub_paths))
+            exp_sub_path = exp_sub_paths[0].content
         else:
-            selected_experiences = None
+            exp_sub_path = ""
+        selected_experiences = self.load_experience(query=query, sub_path=exp_sub_path)
+
+        planning_enrichments = memory.get_shared_memory_entries(entry_type="plan")
+
+        prompt = self.compose_prompt(
+            rounds,
+            self.plugin_pool,
+            selected_experiences,
+            planning_enrichments=[pe.content for pe in planning_enrichments],
+        )
 
-        prompt = self.compose_prompt(rounds, self.plugin_pool, selected_experiences)
         self.tracing.set_span_attribute("prompt", json.dumps(prompt, indent=2))
         prompt_size = self.tracing.count_tokens(json.dumps(prompt))
         self.tracing.set_span_attribute("prompt_size", prompt_size)

diff --git a/taskweaver/code_interpreter/code_interpreter/code_generator_prompt.yaml b/taskweaver/code_interpreter/code_interpreter/code_generator_prompt.yaml
@@ -80,10 +80,10 @@ conversation_head: |-
 
 user_message_head: |-
     -----------------------------
-    # Feedback of the code in the last round (None if no feedback):
+    ### Feedback of the code in the last round (None if no feedback):
     {FEEDBACK}
     
-    # Additional information from the User in this round:
+    ### Request from the User in this round:
     {MESSAGE}
 
 requirements: |-
@@ -96,7 +96,7 @@ requirements: |-
     {CODE_GENERATION_REQUIREMENTS}
 
 experience_instruction: |-
-    ## Experience And Lessons
-    Before generating Python code, please refer to the experiences and lessons learned from the previous tasks:
+    ### Experience And Lessons
+    Before generating code, please learn from the following past experiences and lessons:
     {experiences}
-    You must use the experiences and lessons learned to generate the Python code.
+    You must apply them in code generation.
diff --git a/taskweaver/config/module_config.py b/taskweaver/config/module_config.py
@@ -1,10 +1,11 @@
 from typing import Any, List, Optional
 
-from injector import inject
+from injector import inject, singleton
 
 from taskweaver.config.config_mgt import AppConfigSource
 
 
+@singleton
 class ModuleConfig(object):
     @inject
     def __init__(self, src: AppConfigSource) -> None:

diff --git a/taskweaver/memory/__init__.py b/taskweaver/memory/__init__.py
@@ -1,8 +1,10 @@
 from __future__ import annotations
 
+from .shared_memory_entry import SharedMemoryEntry
 from .attachment import Attachment
 from .conversation import Conversation
 from .memory import Memory
 from .post import Post
 from .round import Round
 from .compression import RoundCompressor
+
diff --git a/taskweaver/memory/attachment.py b/taskweaver/memory/attachment.py
@@ -43,8 +43,8 @@ class AttachmentType(Enum):
     invalid_response = "invalid_response"
     text = "text"
 
-    # board info
-    board = "board"
+    # shared memory entry
+    shared_memory_entry = "shared_memory_entry"
 
 
 @dataclass
@@ -95,11 +95,15 @@ def __str__(self) -> str:
         return self.__repr__()
 
     def to_dict(self) -> AttachmentDict:
+        if self.extra is not None and hasattr(self.extra, "to_dict"):
+            extra_content = self.extra.to_dict()
+        else:
+            extra_content = self.extra
         return {
             "id": self.id,
             "type": self.type.value,
             "content": self.content,
-            "extra": self.extra,
+            "extra": extra_content,
         }
 
     @staticmethod