langroid 0.20.1__tar.gz → 0.22.0__tar.gz

{langroid-0.20.1 → langroid-0.22.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.20.1
+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)
@@ -248,6 +248,12 @@ teacher_task.run()
 <details>
 <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:**
   - **[0.18.0]** [LLMConfig.async_stream_quiet](https://langroid.github.io/langroid/notes/async-streaming/) flag to 
     turn off LLM output in async + stream mode.
@@ -361,7 +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):**
@@ -645,7 +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-0.20.1 → langroid-0.22.0}/README.md

@@ -94,7 +94,7 @@ 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)
@@ -135,6 +135,12 @@ teacher_task.run()
 <details>
 <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:**
   - **[0.18.0]** [LLMConfig.async_stream_quiet](https://langroid.github.io/langroid/notes/async-streaming/) flag to 
     turn off LLM output in async + stream mode.
@@ -248,7 +254,7 @@ 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):**
@@ -532,7 +538,7 @@ 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-0.20.1 → langroid-0.22.0}/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-0.20.1 → langroid-0.22.0}/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-0.20.1 → langroid-0.22.0}/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
@@ -118,6 +117,7 @@ class ArangoChatAgent(ChatAgent):
     def init_state(self) -> None:
         super().init_state()
         self.current_retrieval_aql_query: str = ""
+        self.current_schema_params: ArangoSchemaTool = ArangoSchemaTool()
         self.num_tries = 0  # how many attempts to answer user question
 
     def user_response(
@@ -125,6 +125,8 @@ class ArangoChatAgent(ChatAgent):
         msg: Optional[str | ChatDocument] = None,
     ) -> Optional[ChatDocument]:
         response = super().user_response(msg)
+        if response is None:
+            return None
         response_str = response.content if response is not None else ""
         if response_str != "":
             self.num_tries = 0  # reset number of tries if user responds
@@ -165,7 +167,20 @@ class ArangoChatAgent(ChatAgent):
                 """
             )
 
-        return super().llm_response(message)
+        response = super().llm_response(message)
+        if (
+            response is not None
+            and self.config.chat_mode
+            and self.config.addressing_prefix in response.content
+            and self.has_tool_message_attempt(response)
+        ):
+            # response contains both a user-addressing and a tool, which
+            # is not allowed, so remove the user-addressing prefix
+            response.content = response.content.replace(
+                self.config.addressing_prefix, ""
+            )
+
+        return response
 
     def _validate_config(self) -> None:
         assert isinstance(self.config, ArangoChatAgentConfig)
@@ -327,28 +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:
@@ -363,6 +356,11 @@ class ArangoChatAgent(ChatAgent):
             """
         self.num_tries += 1
         query = msg.aql_query
+        if query == self.current_retrieval_aql_query:
+            return """
+            You have already tried this query, so you will get the same results again!
+            If you need to retry, please MODIFY the query to get different results.
+            """
         self.current_retrieval_aql_query = query
         logger.info(f"Executing AQL query: {query}")
         response = self.read_query(query)
@@ -374,9 +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"""
@@ -399,6 +395,17 @@ class ArangoChatAgent(ChatAgent):
         else show all properties and example-docs.
         """
 
+        if (
+            msg is not None
+            and msg.collections == self.current_schema_params.collections
+            and msg.properties == self.current_schema_params.properties
+        ):
+            return """
+            You have already tried this schema TOOL, so you will get the same results 
+            again! Please MODIFY the tool params `collections` or `properties` to get
+            different results.
+            """
+
         if msg is not None:
             collections = msg.collections
             properties = msg.properties

{langroid-0.20.1 → langroid-0.22.0}/langroid/agent/special/arangodb/system_messages.py

@@ -144,14 +144,17 @@ THEN make ANOTHER query, and so on, until you have the answer.
 RETRY-SUGGESTIONS:
 If you receive a null or other unexpected result,
 (a) make sure you use the available TOOLs correctly,
-(b) USE `{arango_schema_tool_name}` tool/function-call to get all collections, 
-    their attributes and graph definitions available in your ArangoDB database.
+(b) learn more about the schema using EITHER:
+ - `{arango_schema_tool_name}` tool/function-call to find properties of specific
+    collections or other parts of the schema, OR
+ - `{aql_retrieval_tool_name}` tool/function-call to use AQL queries to 
+    find specific parts of the schema.
 (c) Collection names are CASE-SENSITIVE -- make sure you adhere to the exact 
     collection name you found in the schema.
 (d) see if you have made an assumption in your AQL query, and try another way, 
     or use `{aql_retrieval_tool_name}` to explore the database contents before 
     submitting your final query. 
-(f) Try APPROXIMATE or PARTIAL MATCHES to strings in the user's query, 
+(e) Try APPROXIMATE or PARTIAL MATCHES to strings in the user's query, 
     e.g. user may ask about "Godfather" instead of "The Godfather",
     or try using CASE-INSENSITIVE MATCHES.
     

{langroid-0.20.1 → langroid-0.22.0}/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-0.20.1 → langroid-0.22.0}/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-0.20.1 → langroid-0.22.0}/langroid/language_models/__init__.py

@@ -17,6 +17,7 @@ from .base import (
 from .openai_gpt import (
     OpenAIChatModel,
     AnthropicModel,
+    GeminiModel,
     OpenAICompletionModel,
     OpenAIGPTConfig,
     OpenAIGPT,
@@ -41,6 +42,7 @@ __all__ = [
     "LLMResponse",
     "OpenAIChatModel",
     "AnthropicModel",
+    "GeminiModel",
     "OpenAICompletionModel",
     "OpenAIGPTConfig",
     "OpenAIGPT",

{langroid-0.20.1 → langroid-0.22.0}/langroid/language_models/openai_gpt.py

@@ -66,6 +66,7 @@ if "OLLAMA_HOST" in os.environ:
 else:
     OLLAMA_BASE_URL = "http://localhost:11434/v1"
 
+GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta/"
 OLLAMA_API_KEY = "ollama"
 DUMMY_API_KEY = "xxx"
 
@@ -92,6 +93,14 @@ class OpenAIChatModel(str, Enum):
     O1_MINI = "o1-mini"
 
 
+class GeminiModel(str, Enum):
+    """Enum for Gemini models"""
+
+    GEMINI_1_5_FLASH = "gemini-1.5-flash"
+    GEMINI_1_5_FLASH_8B = "gemini-1.5-flash-8b"
+    GEMINI_1_5_PRO = "gemini-1.5-pro"
+
+
 class OpenAICompletionModel(str, Enum):
     """Enum for OpenAI Completion models"""
 
@@ -503,6 +512,7 @@ class OpenAIGPT(LanguageModel):
 
         self.is_groq = self.config.chat_model.startswith("groq/")
         self.is_cerebras = self.config.chat_model.startswith("cerebras/")
+        self.is_gemini = self.config.chat_model.startswith("gemini/")
 
         if self.is_groq:
             self.config.chat_model = self.config.chat_model.replace("groq/", "")
@@ -524,6 +534,11 @@ class OpenAIGPT(LanguageModel):
                 api_key=self.api_key,
             )
         else:
+            if self.is_gemini:
+                self.config.chat_model = self.config.chat_model.replace("gemini/", "")
+                self.api_key = os.getenv("GEMINI_API_KEY", DUMMY_API_KEY)
+                self.api_base = GEMINI_BASE_URL
+
             self.client = OpenAI(
                 api_key=self.api_key,
                 base_url=self.api_base,
@@ -623,7 +638,13 @@ class OpenAIGPT(LanguageModel):
         Currently main troublemaker is o1* series.
         """
         match self.config.chat_model:
-            case OpenAIChatModel.O1_MINI | OpenAIChatModel.O1_PREVIEW:
+            case (
+                OpenAIChatModel.O1_MINI
+                | OpenAIChatModel.O1_PREVIEW
+                | GeminiModel.GEMINI_1_5_FLASH
+                | GeminiModel.GEMINI_1_5_FLASH_8B
+                | GeminiModel.GEMINI_1_5_PRO
+            ):
                 return {"max_tokens": "max_completion_tokens"}
             case _:
                 return {}
@@ -1229,13 +1250,10 @@ class OpenAIGPT(LanguageModel):
         cost = 0.0
         prompt_tokens = 0
         completion_tokens = 0
-        if not cached and not self.get_stream():
+        if not cached and not self.get_stream() and response["usage"] is not None:
             prompt_tokens = response["usage"]["prompt_tokens"]
             completion_tokens = response["usage"]["completion_tokens"]
-            cost = self._cost_chat_model(
-                response["usage"]["prompt_tokens"],
-                response["usage"]["completion_tokens"],
-            )
+            cost = self._cost_chat_model(prompt_tokens, completion_tokens)
 
         return LLMTokenUsage(
             prompt_tokens=prompt_tokens, completion_tokens=completion_tokens, cost=cost

{langroid-0.20.1 → langroid-0.22.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.20.1"
+version = "0.22.0"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"