langroid 0.34.1__tar.gz → 0.35.1__tar.gz

{langroid-0.34.1 → langroid-0.35.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langroid
-Version: 0.34.1
+Version: 0.35.1
 Summary: Harness LLMs with Multi-Agent Programming
 Author-email: Prasad Chalasani <pchalasani@gmail.com>
 License: MIT

{langroid-0.34.1 → langroid-0.35.1}/langroid/agent/chat_agent.py

@@ -416,7 +416,7 @@ class ChatAgent(Agent):
         ]
 
         if len(usable_tool_classes) == 0:
-            return "You can ask questions in natural language."
+            return ""
         format_instructions = "\n\n".join(
             [
                 msg_cls.format_instructions(tool=self.config.use_tools)
@@ -568,17 +568,14 @@ class ChatAgent(Agent):
         Returns:
             LLMMessage object
         """
-        content = textwrap.dedent(
-            f"""
-            {self.system_message}
-            
-            {self.system_tool_instructions}
-            
-            {self.system_tool_format_instructions}
+        content = self.system_message
+        if self.system_tool_instructions != "":
+            content += "\n\n" + self.system_tool_instructions
+        if self.system_tool_format_instructions != "":
+            content += "\n\n" + self.system_tool_format_instructions
+        if self.output_format_instructions != "":
+            content += "\n\n" + self.output_format_instructions
 
-            {self.output_format_instructions}
-            """
-        )
         # remove leading and trailing newlines and other whitespace
         return LLMMessage(role=Role.SYSTEM, content=content.strip())
 

{langroid-0.34.1 → langroid-0.35.1}/langroid/agent/chat_document.py

@@ -117,6 +117,7 @@ class ChatDocument(Document):
         attachment (None | ChatDocAttachment): Any additional data attached.
     """
 
+    reasoning: str = ""  # reasoning produced by a reasoning LLM
     content_any: Any = None  # to hold arbitrary data returned by responders
     oai_tool_calls: Optional[List[OpenAIToolCall]] = None
     oai_tool_id2result: Optional[OrderedDict[str, str]] = None
@@ -297,6 +298,7 @@ class ChatDocument(Document):
                 ChatDocument._clean_fn_call(oai_tc.function)
         return ChatDocument(
             content=message,
+            reasoning=response.reasoning,
             content_any=message,
             oai_tool_calls=response.oai_tool_calls,
             function_call=response.function_call,

{langroid-0.34.1 → langroid-0.35.1}/langroid/agent/special/doc_chat_agent.py

@@ -212,6 +212,12 @@ class DocChatAgentConfig(ChatAgentConfig):
     )
 
 
+def _append_metadata_source(orig_source: str, source: str) -> str:
+    if orig_source != source and source != "" and orig_source != "":
+        return f"{orig_source.strip()}; {source.strip()}"
+    return orig_source.strip() + source.strip()
+
+
 class DocChatAgent(ChatAgent):
     """
     Agent for chatting with a collection of documents.
@@ -337,7 +343,11 @@ class DocChatAgent(ChatAgent):
                 url_docs = loader.load()
                 # update metadata of each doc with meta
                 for d in url_docs:
+                    orig_source = d.metadata.source
                     d.metadata = d.metadata.copy(update=meta)
+                    d.metadata.source = _append_metadata_source(
+                        orig_source, meta.get("source", "")
+                    )
                 docs.extend(url_docs)
         if len(paths) > 0:  # paths OR bytes are handled similarly
             for pi in path_idxs:
@@ -350,7 +360,11 @@ class DocChatAgent(ChatAgent):
                 )
                 # update metadata of each doc with meta
                 for d in path_docs:
+                    orig_source = d.metadata.source
                     d.metadata = d.metadata.copy(update=meta)
+                    d.metadata.source = _append_metadata_source(
+                        orig_source, meta.get("source", "")
+                    )
                 docs.extend(path_docs)
         n_docs = len(docs)
         n_splits = self.ingest_docs(docs, split=self.config.split)
@@ -391,15 +405,26 @@ class DocChatAgent(ChatAgent):
         """
         if isinstance(metadata, list) and len(metadata) > 0:
             for d, m in zip(docs, metadata):
-                d.metadata = d.metadata.copy(
-                    update=m if isinstance(m, dict) else m.dict()  # type: ignore
+                orig_source = d.metadata.source
+                m_dict = m if isinstance(m, dict) else m.dict()  # type: ignore
+                d.metadata = d.metadata.copy(update=m_dict)  # type: ignore
+                d.metadata.source = _append_metadata_source(
+                    orig_source, m_dict.get("source", "")
                 )
         elif isinstance(metadata, dict):
             for d in docs:
+                orig_source = d.metadata.source
                 d.metadata = d.metadata.copy(update=metadata)
+                d.metadata.source = _append_metadata_source(
+                    orig_source, metadata.get("source", "")
+                )
         elif isinstance(metadata, DocMetaData):
             for d in docs:
+                orig_source = d.metadata.source
                 d.metadata = d.metadata.copy(update=metadata.dict())
+                d.metadata.source = _append_metadata_source(
+                    orig_source, metadata.source
+                )
 
         self.original_docs.extend(docs)
         if self.parser is None:

{langroid-0.34.1 → langroid-0.35.1}/langroid/language_models/base.py

@@ -75,6 +75,8 @@ class LLMConfig(BaseSettings):
     # use chat model for completion? For OpenAI models, this MUST be set to True!
     use_chat_for_completion: bool = True
     stream: bool = True  # stream output from API?
+    # TODO: we could have a `stream_reasoning` flag here to control whether to show
+    # reasoning output from reasoning models
     cache_config: None | CacheDBConfig = RedisCacheConfig()
 
     # Dict of model -> (input/prompt cost, output/completion cost)
@@ -317,6 +319,7 @@ class LLMResponse(BaseModel):
     """
 
     message: str
+    reasoning: str = ""  # optional reasoning text from reasoning models
     # TODO tool_id needs to generalize to multi-tool calls
     tool_id: str = ""  # used by OpenAIAssistant
     oai_tool_calls: Optional[List[OpenAIToolCall]] = None

{langroid-0.34.1 → langroid-0.35.1}/langroid/language_models/openai_gpt.py

@@ -820,6 +820,7 @@ class OpenAIGPT(LanguageModel):
         tool_deltas: List[Dict[str, Any]] = [],
         has_function: bool = False,
         completion: str = "",
+        reasoning: str = "",
         function_args: str = "",
         function_name: str = "",
     ) -> Tuple[bool, bool, str, str]:
@@ -845,7 +846,9 @@ class OpenAIGPT(LanguageModel):
         # In the 1st: choices list is empty, in the 2nd: the dict delta has null content
         if chat:
             delta = choices[0].get("delta", {})
+            # capture both content and reasoning_content
             event_text = delta.get("content", "")
+            event_reasoning = delta.get("reasoning_content", "")
             if "function_call" in delta and delta["function_call"] is not None:
                 if "name" in delta["function_call"]:
                     event_fn_name = delta["function_call"]["name"]
@@ -857,6 +860,7 @@ class OpenAIGPT(LanguageModel):
                 tool_deltas += event_tool_deltas
         else:
             event_text = choices[0]["text"]
+            event_reasoning = ""  # TODO: Ignoring reasoning for non-chat models
 
         finish_reason = choices[0].get("finish_reason", "")
         if not event_text and finish_reason == "content_filter":
@@ -877,6 +881,11 @@ class OpenAIGPT(LanguageModel):
             sys.stdout.write(Colors().GREEN + event_text)
             sys.stdout.flush()
             self.config.streamer(event_text, StreamEventType.TEXT)
+        if event_reasoning:
+            reasoning += event_reasoning
+            sys.stdout.write(Colors().GREEN_DIM + event_reasoning)
+            sys.stdout.flush()
+            self.config.streamer(event_reasoning, StreamEventType.TEXT)
         if event_fn_name:
             function_name = event_fn_name
             has_function = True
@@ -915,8 +924,15 @@ class OpenAIGPT(LanguageModel):
             # for function_call, finish_reason does not necessarily
             # contain "function_call" as mentioned in the docs.
             # So we check for "stop" or "function_call" here.
-            return True, has_function, function_name, function_args, completion
-        return False, has_function, function_name, function_args, completion
+            return (
+                True,
+                has_function,
+                function_name,
+                function_args,
+                completion,
+                reasoning,
+            )
+        return False, has_function, function_name, function_args, completion, reasoning
 
     @no_type_check
     async def _process_stream_event_async(
@@ -926,6 +942,7 @@ class OpenAIGPT(LanguageModel):
         tool_deltas: List[Dict[str, Any]] = [],
         has_function: bool = False,
         completion: str = "",
+        reasoning: str = "",
         function_args: str = "",
         function_name: str = "",
     ) -> Tuple[bool, bool, str, str]:
@@ -953,6 +970,7 @@ class OpenAIGPT(LanguageModel):
         if chat:
             delta = choices[0].get("delta", {})
             event_text = delta.get("content", "")
+            event_reasoning = delta.get("reasoning_content", "")
             if "function_call" in delta and delta["function_call"] is not None:
                 if "name" in delta["function_call"]:
                     event_fn_name = delta["function_call"]["name"]
@@ -964,12 +982,19 @@ class OpenAIGPT(LanguageModel):
                 tool_deltas += event_tool_deltas
         else:
             event_text = choices[0]["text"]
+            event_reasoning = ""  # TODO: Ignoring reasoning for non-chat models
         if event_text:
             completion += event_text
             if not silent:
                 sys.stdout.write(Colors().GREEN + event_text)
                 sys.stdout.flush()
                 await self.config.streamer_async(event_text, StreamEventType.TEXT)
+        if event_reasoning:
+            reasoning += event_reasoning
+            if not silent:
+                sys.stdout.write(Colors().GREEN + event_reasoning)
+                sys.stdout.flush()
+                await self.config.streamer_async(event_reasoning, StreamEventType.TEXT)
         if event_fn_name:
             function_name = event_fn_name
             has_function = True
@@ -1016,8 +1041,15 @@ class OpenAIGPT(LanguageModel):
             # for function_call, finish_reason does not necessarily
             # contain "function_call" as mentioned in the docs.
             # So we check for "stop" or "function_call" here.
-            return True, has_function, function_name, function_args, completion
-        return False, has_function, function_name, function_args, completion
+            return (
+                True,
+                has_function,
+                function_name,
+                function_args,
+                completion,
+                reasoning,
+            )
+        return False, has_function, function_name, function_args, completion, reasoning
 
     @retry_with_exponential_backoff
     def _stream_response(  # type: ignore
@@ -1035,6 +1067,7 @@ class OpenAIGPT(LanguageModel):
 
         """
         completion = ""
+        reasoning = ""
         function_args = ""
         function_name = ""
 
@@ -1050,12 +1083,14 @@ class OpenAIGPT(LanguageModel):
                     function_name,
                     function_args,
                     completion,
+                    reasoning,
                 ) = self._process_stream_event(
                     event,
                     chat=chat,
                     tool_deltas=tool_deltas,
                     has_function=has_function,
                     completion=completion,
+                    reasoning=reasoning,
                     function_args=function_args,
                     function_name=function_name,
                 )
@@ -1072,6 +1107,7 @@ class OpenAIGPT(LanguageModel):
             tool_deltas=tool_deltas,
             has_function=has_function,
             completion=completion,
+            reasoning=reasoning,
             function_args=function_args,
             function_name=function_name,
         )
@@ -1092,6 +1128,7 @@ class OpenAIGPT(LanguageModel):
 
         """
         completion = ""
+        reasoning = ""
         function_args = ""
         function_name = ""
 
@@ -1107,12 +1144,14 @@ class OpenAIGPT(LanguageModel):
                     function_name,
                     function_args,
                     completion,
+                    reasoning,
                 ) = await self._process_stream_event_async(
                     event,
                     chat=chat,
                     tool_deltas=tool_deltas,
                     has_function=has_function,
                     completion=completion,
+                    reasoning=reasoning,
                     function_args=function_args,
                     function_name=function_name,
                 )
@@ -1129,6 +1168,7 @@ class OpenAIGPT(LanguageModel):
             tool_deltas=tool_deltas,
             has_function=has_function,
             completion=completion,
+            reasoning=reasoning,
             function_args=function_args,
             function_name=function_name,
         )
@@ -1263,6 +1303,7 @@ class OpenAIGPT(LanguageModel):
         tool_deltas: List[Dict[str, Any]] = [],
         has_function: bool = False,
         completion: str = "",
+        reasoning: str = "",
         function_args: str = "",
         function_name: str = "",
     ) -> Tuple[LLMResponse, Dict[str, Any]]:
@@ -1298,7 +1339,12 @@ class OpenAIGPT(LanguageModel):
                 tool_deltas,
             )
             completion = completion + "\n" + failed_content
-            msg: Dict[str, Any] = dict(message=dict(content=completion))
+            msg: Dict[str, Any] = dict(
+                message=dict(
+                    content=completion,
+                    reasoning_content=reasoning,
+                ),
+            )
             if len(tool_dicts) > 0:
                 msg["message"]["tool_calls"] = tool_dicts
 
@@ -1314,6 +1360,7 @@ class OpenAIGPT(LanguageModel):
         else:
             # non-chat mode has no function_call
             msg = dict(text=completion)
+            # TODO: Ignoring reasoning content for non-chat models
 
         # create an OpenAIResponse object so we can cache it as if it were
         # a non-streaming response
@@ -1324,6 +1371,7 @@ class OpenAIGPT(LanguageModel):
         return (
             LLMResponse(
                 message=completion,
+                reasoning=reasoning,
                 cached=False,
                 # don't allow empty list [] here
                 oai_tool_calls=tool_calls or None if len(tool_deltas) > 0 else None,
@@ -1461,6 +1509,7 @@ class OpenAIGPT(LanguageModel):
         )
         args = self._openai_api_call_params(args)
         cached, hashed_key, response = completions_with_backoff(**args)
+        # assume response is an actual response rather than a streaming event
         if not isinstance(response, dict):
             response = response.dict()
         if "message" in response["choices"][0]:
@@ -1534,6 +1583,7 @@ class OpenAIGPT(LanguageModel):
             max_tokens=max_tokens,
             stream=False,
         )
+        # assume response is an actual response rather than a streaming event
         if not isinstance(response, dict):
             response = response.dict()
         if "message" in response["choices"][0]:
@@ -1738,7 +1788,9 @@ class OpenAIGPT(LanguageModel):
             if (
                 len(llm_messages) == 1
                 and llm_messages[0].role == Role.SYSTEM
-                and self.requires_first_user_message()
+                # TODO: we will unconditionally insert a dummy user msg
+                # if the only msg is a system msg.
+                # and self.requires_first_user_message()
             ):
                 # some LLMs, notable Gemini as of 12/11/24,
                 # require the first message to be from the user,
@@ -1827,6 +1879,7 @@ class OpenAIGPT(LanguageModel):
                     "role": "assistant",
                     "name": "",
                     "content": "\n\nHello there, how may I help you?",
+                    "reasoning_content": "Okay, let's see here, hmmm...",
                     "function_call": {
                         "name": "fun_name",
                         "arguments: {
@@ -1845,7 +1898,8 @@ class OpenAIGPT(LanguageModel):
         }
         """
         message = response["choices"][0]["message"]
-        msg = message["content"] or ""
+        msg = message.get("content", "")
+        reasoning = message.get("reasoning_content", "")
 
         if message.get("function_call") is None:
             fun_call = None
@@ -1879,6 +1933,7 @@ class OpenAIGPT(LanguageModel):
                     msg = msg + "\n" + json.dumps(tool_call_dict)
         return LLMResponse(
             message=msg.strip() if msg is not None else "",
+            reasoning=reasoning.strip() if reasoning is not None else "",
             function_call=fun_call,
             oai_tool_calls=oai_tool_calls or None,  # don't allow empty list [] here
             cached=cached,

{langroid-0.34.1 → langroid-0.35.1}/langroid/utils/constants.py

@@ -6,6 +6,8 @@ class Colors(BaseModel):
     RED: str = "\033[31m"
     BLUE: str = "\033[34m"
     GREEN: str = "\033[32m"
+    GREEN_DIMMER: str = "\033[38;5;22m"  # very dark green
+    GREEN_DIM: str = "\033[38;5;28m"  # medium-dim green
     ORANGE: str = "\033[33m"  # no standard ANSI color for orange; using yellow
     CYAN: str = "\033[36m"
     MAGENTA: str = "\033[35m"

{langroid-0.34.1 → langroid-0.35.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "langroid"
-version = "0.34.1"
+version = "0.35.1"
 authors = [
     {name = "Prasad Chalasani", email = "pchalasani@gmail.com"},
 ]
@@ -220,13 +220,13 @@ dev = [
     "pytest-redis<4.0.0,>=3.0.2",
     "pytest-asyncio<1.0.0,>=0.21.1",
     "pytest-postgresql<6.0.0,>=5.0.0",
-    "pytest-mysql<3.0.0,>=2.4.2",
     "coverage<8.0.0,>=7.2.5",
     "pytest-xdist<4.0.0,>=3.6.1",
     "pytest-timeout<3.0.0,>=2.3.1",
     "pytest-cov<6.0.0,>=5.0.0",
     "docker<8.0.0,>=7.1.0",
     "commitizen>=4.1.0",
+    "pytest-mysql>=3.1.0",
 ]
 docs = [
     "mkdocs<2.0.0,>=1.4.2",