langroid 0.21.0__py3-none-any.whl → 0.22.0__py3-none-any.whl

langroid/agent/base.py

@@ -1450,6 +1450,40 @@ class Agent(ABC):
 
         return None
 
+    def _maybe_truncate_result(
+        self, result: str | ChatDocument | None, max_tokens: int | None
+    ) -> str | ChatDocument | None:
+        """
+        Truncate the result string to `max_tokens` tokens.
+        """
+        if result is None or max_tokens is None:
+            return result
+        result_str = result.content if isinstance(result, ChatDocument) else result
+        num_tokens = (
+            self.parser.num_tokens(result_str)
+            if self.parser is not None
+            else len(result_str) / 4.0
+        )
+        if num_tokens <= max_tokens:
+            return result
+        truncate_warning = f"""
+        The TOOL result was large, so it was truncated to {max_tokens} tokens.
+        To get the full result, the TOOL must be called again.
+        """
+        if isinstance(result, str):
+            return (
+                self.parser.truncate_tokens(result, max_tokens)
+                if self.parser is not None
+                else result[: max_tokens * 4]  # approx truncate
+            ) + truncate_warning
+        elif isinstance(result, ChatDocument):
+            result.content = (
+                self.parser.truncate_tokens(result.content, max_tokens)
+                if self.parser is not None
+                else result.content[: max_tokens * 4]  # approx truncate
+            ) + truncate_warning
+            return result
+
     def handle_tool_message(
         self,
         tool: ToolMessage,
@@ -1485,7 +1519,9 @@ class Agent(ABC):
             # not a pydantic validation error,
             # which we check in `handle_message`
             raise e
-        return result  # type: ignore
+        return self._maybe_truncate_result(
+            result, tool._max_result_tokens
+        )  # type: ignore
 
     def num_tokens(self, prompt: str | List[LLMMessage]) -> int:
         if self.parser is None:

langroid/agent/chat_agent.py

@@ -623,6 +623,33 @@ class ChatAgent(Agent):
             self.llm_tools_usable.discard(r)
             self.llm_functions_usable.discard(r)
 
+    def _reduce_raw_tool_results(self, message: ChatDocument) -> None:
+        """
+        If message is the result of a ToolMessage that had the
+        flag `_retain_raw_results = False`, then we replace contents
+        with a placeholder message.
+        """
+        parent_message: ChatDocument | None = message.parent
+        tools = [] if parent_message is None else parent_message.tool_messages
+        truncate_tools = [t for t in tools if t._max_retained_tokens is not None]
+        limiting_tool = truncate_tools[0] if len(truncate_tools) > 0 else None
+        if limiting_tool is not None and limiting_tool._max_retained_tokens is not None:
+            tool_name = limiting_tool.default_value("request")
+            max_tokens: int = limiting_tool._max_retained_tokens
+            truncation_warning = f"""
+                The result of the {tool_name} tool were too large, 
+                and has been truncated to {max_tokens} tokens.
+                To obtain the full result, the tool needs to be re-used.
+            """
+            llm_msg = self.message_history[message.metadata.msg_idx]
+            orig_content = llm_msg.content
+            new_content = (
+                self.parser.truncate_tokens(orig_content, max_tokens)
+                if self.parser is not None
+                else orig_content[: max_tokens * 4]  # approx truncation
+            )
+            llm_msg.content = new_content + "\n\n" + truncation_warning
+
     def llm_response(
         self, message: Optional[str | ChatDocument] = None
     ) -> Optional[ChatDocument]:
@@ -650,6 +677,8 @@ class ChatAgent(Agent):
         self.message_history.extend(ChatDocument.to_LLMMessage(response))
         response.metadata.msg_idx = len(self.message_history) - 1
         response.metadata.agent_id = self.id
+        if isinstance(message, ChatDocument):
+            self._reduce_raw_tool_results(message)
         # Preserve trail of tool_ids for OpenAI Assistant fn-calls
         response.metadata.tool_ids = (
             []
@@ -681,6 +710,8 @@ class ChatAgent(Agent):
         self.message_history.extend(ChatDocument.to_LLMMessage(response))
         response.metadata.msg_idx = len(self.message_history) - 1
         response.metadata.agent_id = self.id
+        if isinstance(message, ChatDocument):
+            self._reduce_raw_tool_results(message)
         # Preserve trail of tool_ids for OpenAI Assistant fn-calls
         response.metadata.tool_ids = (
             []

langroid/agent/special/arangodb/arangodb_agent.py

@@ -94,7 +94,6 @@ class ArangoChatAgentConfig(ChatAgentConfig):
     prepopulate_schema: bool = True
     use_functions_api: bool = True
     max_num_results: int = 10  # how many results to return from AQL query
-    max_result_tokens: int = 1000  # truncate long results to this many tokens
     max_schema_fields: int = 500  # max fields to show in schema
     max_tries: int = 10  # how many attempts to answer user question
     use_tools: bool = False
@@ -343,28 +342,6 @@ class ArangoChatAgent(ChatAgent):
                 success=False, data=f"Failed after max retries: {str(e)}"
             )
 
-    def _limit_tokens(self, text: str) -> str:
-        result = text
-        n_toks = self.num_tokens(result)
-        if n_toks > self.config.max_result_tokens:
-            logger.warning(
-                f"""
-                Your query resulted in a large result of  
-                {n_toks} tokens,
-                which will be truncated to {self.config.max_result_tokens} tokens.
-                If this does not give satisfactory results,
-                please retry with a more focused query.
-                """
-            )
-            if self.parser is not None:
-                result = self.parser.truncate_tokens(
-                    result,
-                    self.config.max_result_tokens,
-                )
-            else:
-                result = result[: self.config.max_result_tokens * 4]  # truncate roughly
-        return result
-
     def aql_retrieval_tool(self, msg: AQLRetrievalTool) -> str:
         """Handle AQL query for data retrieval"""
         if not self.tried_schema:
@@ -395,9 +372,7 @@ class ArangoChatAgent(ChatAgent):
             Try modifying your query based on the RETRY-SUGGESTIONS 
             in your instructions.
             """
-        # truncate long results
-        result = str(response.data)
-        return self._limit_tokens(result)
+        return str(response.data)
 
     def aql_creation_tool(self, msg: AQLCreationTool) -> str:
         """Handle AQL query for creating data"""

langroid/agent/special/arangodb/tools.py

@@ -13,6 +13,9 @@ class AQLRetrievalTool(ToolMessage):
     """
     aql_query: str
 
+    _max_result_tokens = 500
+    _max_retained_tokens = 200
+
     @classmethod
     def examples(cls) -> List[ToolMessage | Tuple[str, ToolMessage]]:
         """Few-shot examples to include in tool instructions."""
@@ -98,5 +101,7 @@ class ArangoSchemaTool(ToolMessage):
     properties: bool = True
     collections: List[str] | None = None
 
+    _max_result_tokens = 500
+
 
 arango_schema_tool_name = ArangoSchemaTool.default_value("request")

langroid/agent/tool_message.py

@@ -44,7 +44,19 @@ class ToolMessage(ABC, BaseModel):
 
     _allow_llm_use: bool = True  # allow an LLM to use (i.e. generate) this tool?
 
-    # model_config = ConfigDict(extra=Extra.allow)
+    # Optional param to limit number of result tokens to retain in msg history.
+    # Some tools can have large results that we may not want to fully retain,
+    # e.g. result of a db query, which the LLM later reduces to a summary, so
+    # in subsequent dialog we may only want to retain the summary,
+    # and replace this raw result truncated to _max_result_tokens.
+    # Important to note: unlike _max_result_tokens, this param is used
+    # NOT used to immediately truncate the result;
+    # it is only used to truncate what is retained in msg history AFTER the
+    # response to this result.
+    _max_retained_tokens: int | None = None
+
+    # Optional param to limit number of tokens in the result of the tool.
+    _max_result_tokens: int | None = None
 
     class Config:
         extra = Extra.allow

{langroid-0.21.0.dist-info → langroid-0.22.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.21.0
+Version: 0.22.0
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani
@@ -207,7 +207,7 @@ import langroid.language_models as lm
 # set up LLM
 llm_cfg = lm.OpenAIGPTConfig( # or OpenAIAssistant to use Assistant API 
   # any model served via an OpenAI-compatible API
-  chat_model=lm.OpenAIChatModel.GPT4_TURBO, # or, e.g., "ollama/mistral"
+  chat_model=lm.OpenAIChatModel.GPT4o, # or, e.g., "ollama/mistral"
 )
 # use LLM directly
 mdl = lm.OpenAIGPT(llm_cfg)
@@ -249,6 +249,9 @@ teacher_task.run()
 <summary> <b>Click to expand</b></summary>
 
 - **Nov 2024:**
+  - **[0.22.0](https://langroid.github.io/langroid/notes/large-tool-results/)**: 
+     Optional parameters to truncate large tool results.
+  - **[0.21.0](https://langroid.github.io/langroid/notes/gemini/)** Direct support for Gemini models via OpenAI client instead of using LiteLLM.
   - **[0.20.0](https://github.com/langroid/langroid/releases/tag/0.20.0)** Support for 
     ArangoDB Knowledge Graphs.
 - **Oct 2024:**
@@ -364,7 +367,7 @@ teacher_task.run()
       decides a filter and rephrase query to send to a RAG agent.
   - **[0.1.141](https://github.com/langroid/langroid/releases/tag/0.1.141):**
     API Simplifications to reduce boilerplate:
-    auto-select an available OpenAI model (preferring gpt-4-turbo), simplifies defaults.
+    auto-select an available OpenAI model (preferring gpt-4o), simplifies defaults.
     Simpler `Task` initialization with default `ChatAgent`.
 - **Nov 2023:**
   - **[0.1.126](https://github.com/langroid/langroid/releases/tag/0.1.126):**
@@ -648,7 +651,7 @@ provides more information, and you can set each environment variable as follows:
 - `AZURE_OPENAI_API_BASE` from the value of `ENDPOINT`, typically looks like `https://your.domain.azure.com`.
 - For `AZURE_OPENAI_API_VERSION`, you can use the default value in `.env-template`, and latest version can be found [here](https://learn.microsoft.com/en-us/azure/ai-services/openai/whats-new#azure-openai-chat-completion-general-availability-ga)
 - `AZURE_OPENAI_DEPLOYMENT_NAME` is the name of the deployed model, which is defined by the user during the model setup 
-- `AZURE_OPENAI_MODEL_NAME` Azure OpenAI allows specific model names when you select the model for your deployment. You need to put precisly the exact model name that was selected. For example, GPT-3.5 (should be `gpt-35-turbo-16k` or `gpt-35-turbo`) or GPT-4 (should be `gpt-4-32k` or `gpt-4`).
+- `AZURE_OPENAI_MODEL_NAME` Azure OpenAI allows specific model names when you select the model for your deployment. You need to put precisly the exact model name that was selected. For example, GPT-4 (should be `gpt-4-32k` or `gpt-4`).
 - `AZURE_OPENAI_MODEL_VERSION` is required if `AZURE_OPENAI_MODEL_NAME = gpt=4`, which will assist Langroid to determine the cost of the model  
 </details>
 

{langroid-0.21.0.dist-info → langroid-0.22.0.dist-info}/RECORD

@@ -1,19 +1,19 @@
 langroid/__init__.py,sha256=z_fCOLQJPOw3LLRPBlFB5-2HyCjpPgQa4m4iY5Fvb8Y,1800
 langroid/agent/__init__.py,sha256=ll0Cubd2DZ-fsCMl7e10hf9ZjFGKzphfBco396IKITY,786
-langroid/agent/base.py,sha256=sOZapdzHaB4kbCLu8vI_zZx78jIhv9fmWn0EWV4yTAE,65371
+langroid/agent/base.py,sha256=FEAIVwFKlmkOSV3RfRR0-v5UofJBfsBUUSsSEE2ioUI,66824
 langroid/agent/batch.py,sha256=QZdlt1563hx4l3AXrCaGovE-PNG93M3DsvQAbDzdiS8,13705
 langroid/agent/callbacks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 langroid/agent/callbacks/chainlit.py,sha256=JJXI3UGTyTDg2FFath4rqY1GyUo_0pbVBt8CZpvdtn4,23289
-langroid/agent/chat_agent.py,sha256=GVuKXAwHACeTc1gC0y7Ywj8CmJfqu7InJ0AXtmz8nRw,50000
+langroid/agent/chat_agent.py,sha256=2aUDqM6jsdeU8xt-P86JU2TuZhTpPkUVx9mkXFMy8Rs,51646
 langroid/agent/chat_document.py,sha256=FZ_PkeKU5OVp1IUlMvspfqxIXzlyd7J_F32DSYrxQ7E,17651
 langroid/agent/helpers.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 langroid/agent/junk,sha256=LxfuuW7Cijsg0szAzT81OjWWv1PMNI-6w_-DspVIO2s,339
 langroid/agent/openai_assistant.py,sha256=2rjCZw45ysNBEGNzQM4uf0bTC4KkatGYAWcVcW4xcek,34337
 langroid/agent/special/__init__.py,sha256=gik_Xtm_zV7U9s30Mn8UX3Gyuy4jTjQe9zjiE3HWmEo,1273
 langroid/agent/special/arangodb/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-langroid/agent/special/arangodb/arangodb_agent.py,sha256=CXD1Z4yXnWwre4ybg7BMhVFTgQq0zVXzobutyxIYCP4,26862
+langroid/agent/special/arangodb/arangodb_agent.py,sha256=12Y54c84c9qXV-YXRBcI5HaqyiY75JR4TmqlURYKJAM,25851
 langroid/agent/special/arangodb/system_messages.py,sha256=udwfLleTdyz_DuxHuoiv2wHEZoAPBPbwdF_ivjIfP5c,6867
-langroid/agent/special/arangodb/tools.py,sha256=WasFERC1cToLOWi1cWqUs-TujU0A68gZWbhbP128obo,3499
+langroid/agent/special/arangodb/tools.py,sha256=Mixl9WS0r0Crd4nrw2YAB0eY33fTsKISul1053eyeio,3590
 langroid/agent/special/arangodb/utils.py,sha256=LIevtkayIdVVXyj3jlbKH2WgdZTtH5-JLgbXOHC7uxs,1420
 langroid/agent/special/doc_chat_agent.py,sha256=xIqBOyLax_jMU0UevxqXf_aQUrRkW6MQUKpKnKvaqkQ,59281
 langroid/agent/special/lance_doc_chat_agent.py,sha256=s8xoRs0gGaFtDYFUSIRchsgDVbS5Q3C2b2mr3V1Fd-Q,10419
@@ -39,7 +39,7 @@ langroid/agent/special/sql/utils/tools.py,sha256=vFYysk6Vi7HJjII8B4RitA3pt_z3gkS
 langroid/agent/special/table_chat_agent.py,sha256=d9v2wsblaRx7oMnKhLV7uO_ujvk9gh59pSGvBXyeyNc,9659
 langroid/agent/structured_message.py,sha256=y7pud1EgRNeTFZlJmBkLmwME3yQJ_IYik-Xds9kdZbY,282
 langroid/agent/task.py,sha256=f7clh6p6Md0G4YGHqbFeeT88U4XoP0i3eatekV21hHE,86643
-langroid/agent/tool_message.py,sha256=jkN7uq7YwUC_wBcSCNUYjrB_His2YCfQay_lqIa4Tww,10498
+langroid/agent/tool_message.py,sha256=qQebCWogYTtptznIF1mLERwspWjmVxDWFJeERBO-YsI,11207
 langroid/agent/tools/__init__.py,sha256=IMgCte-_ZIvCkozGQmvMqxIw7_nKLKzD78ccJL1bnQU,804
 langroid/agent/tools/duckduckgo_search_tool.py,sha256=NhsCaGZkdv28nja7yveAhSK_w6l_Ftym8agbrdzqgfo,1935
 langroid/agent/tools/file_tools.py,sha256=GjPB5YDILucYapElnvvoYpGJuZQ25ecLs2REv7edPEo,7292
@@ -142,8 +142,8 @@ langroid/vector_store/meilisearch.py,sha256=6frB7GFWeWmeKzRfLZIvzRjllniZ1cYj3Hmh
 langroid/vector_store/momento.py,sha256=qR-zBF1RKVHQZPZQYW_7g-XpTwr46p8HJuYPCkfJbM4,10534
 langroid/vector_store/qdrant_cloud.py,sha256=3im4Mip0QXLkR6wiqVsjV1QvhSElfxdFSuDKddBDQ-4,188
 langroid/vector_store/qdrantdb.py,sha256=v88lqFkepADvlN6lByUj9I4NEKa9X9lWH16uTPPbYrE,17457
-pyproject.toml,sha256=yiLcLTge7QxJ1EWYE05yUHpBADouFh36WO3SlRzJKL0,7488
-langroid-0.21.0.dist-info/LICENSE,sha256=EgVbvA6VSYgUlvC3RvPKehSg7MFaxWDsFuzLOsPPfJg,1065
-langroid-0.21.0.dist-info/METADATA,sha256=TSl0anzCbcnchyI2yTczV4fK2KBkfJnUu2Jl8oqGGLA,56893
-langroid-0.21.0.dist-info/WHEEL,sha256=Nq82e9rUAnEjt98J6MlVmMCZb-t9cYE2Ir1kpBmnWfs,88
-langroid-0.21.0.dist-info/RECORD,,
+pyproject.toml,sha256=uLe37eP_bS9-a9g2l1uNFzDJgpCRqJvbmmwmLQlD3AA,7488
+langroid-0.22.0.dist-info/LICENSE,sha256=EgVbvA6VSYgUlvC3RvPKehSg7MFaxWDsFuzLOsPPfJg,1065
+langroid-0.22.0.dist-info/METADATA,sha256=M15Qy48jMSppZvJRZLNNVLMZpQR9gNoDeFjoLBCKr6M,57107
+langroid-0.22.0.dist-info/WHEEL,sha256=Nq82e9rUAnEjt98J6MlVmMCZb-t9cYE2Ir1kpBmnWfs,88
+langroid-0.22.0.dist-info/RECORD,,

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.21.0"
+version = "0.22.0"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"