langroid 0.49.0__tar.gz → 0.50.0__tar.gz

{langroid-0.49.0 → langroid-0.50.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langroid
-Version: 0.49.0
+Version: 0.50.0
 Summary: Harness LLMs with Multi-Agent Programming
 Author-email: Prasad Chalasani <pchalasani@gmail.com>
 License: MIT
@@ -27,6 +27,7 @@ Requires-Dist: halo<1.0.0,>=0.0.31
 Requires-Dist: jinja2<4.0.0,>=3.1.2
 Requires-Dist: json-repair<1.0.0,>=0.29.9
 Requires-Dist: lxml<5.0.0,>=4.9.3
+Requires-Dist: markdownify>=0.13.1
 Requires-Dist: nest-asyncio<2.0.0,>=1.6.0
 Requires-Dist: nltk<4.0.0,>=3.8.2
 Requires-Dist: onnxruntime<2.0.0,>=1.16.1

{langroid-0.49.0 → langroid-0.50.0}/langroid/agent/special/doc_chat_agent.py

@@ -174,7 +174,7 @@ class DocChatAgentConfig(ChatAgentConfig):
         "https://ai.googleblog.com/2022/11/characterizing-emergent-phenomena-in.html",
     ]
     parsing: ParsingConfig = ParsingConfig(  # modify as needed
-        splitter=Splitter.TOKENS,
+        splitter=Splitter.MARKDOWN,
         chunk_size=1000,  # aim for this many tokens per chunk
         overlap=100,  # overlap between chunks
         max_chunks=10_000,

{langroid-0.49.0 → langroid-0.50.0}/langroid/language_models/openai_gpt.py

@@ -1059,8 +1059,8 @@ class OpenAIGPT(LanguageModel):
                 )
                 if is_break:
                     break
-        except Exception:
-            pass
+        except Exception as e:
+            logging.warning("Error while processing stream response: %s", str(e))
 
         print("")
         # TODO- get usage info in stream mode (?)
@@ -1121,8 +1121,8 @@ class OpenAIGPT(LanguageModel):
                 )
                 if is_break:
                     break
-        except Exception:
-            pass
+        except Exception as e:
+            logging.warning("Error while processing stream response: %s", str(e))
 
         print("")
         # TODO- get usage info in stream mode (?)
@@ -1703,11 +1703,32 @@ class OpenAIGPT(LanguageModel):
             if self.config.litellm and settings.debug:
                 kwargs["logger_fn"] = litellm_logging_fn
             result = completion_call(**kwargs)
-            if not self.get_stream():
-                # if streaming, cannot cache result
+
+            if self.get_stream():
+                # If streaming, cannot cache result
                 # since it is a generator. Instead,
                 # we hold on to the hashed_key and
                 # cache the result later
+
+                # Test if this is a stream with an exception by
+                # trying to get first chunk: Some providers like LiteLLM
+                # produce a valid stream object `result` instead of throwing a
+                # rate-limit error, and if we don't catch it here,
+                # we end up returning an empty response and not
+                # using the retry mechanism in the decorator.
+                try:
+                    # try to get the first chunk to check for errors
+                    test_iter = iter(result)
+                    first_chunk = next(test_iter)
+                    # If we get here without error, recreate the stream
+                    result = chain([first_chunk], test_iter)
+                except StopIteration:
+                    # Empty stream is fine
+                    pass
+                except Exception as e:
+                    # Propagate any errors in the stream
+                    raise e
+            else:
                 self._cache_store(hashed_key, result.model_dump())
         return cached, hashed_key, result
 
@@ -1734,7 +1755,35 @@ class OpenAIGPT(LanguageModel):
                 kwargs["logger_fn"] = litellm_logging_fn
             # If it's not in the cache, call the API
             result = await acompletion_call(**kwargs)
-            if not self.get_stream():
+            if self.get_stream():
+                try:
+                    # Try to peek at the first chunk to immediately catch any errors
+                    # Store the original result (the stream)
+                    original_stream = result
+
+                    # Manually create and advance the iterator to check for errors
+                    stream_iter = original_stream.__aiter__()
+                    try:
+                        # This will raise an exception if the stream is invalid
+                        first_chunk = await anext(stream_iter)
+
+                        # If we reach here, the stream started successfully
+                        # Now recreate a fresh stream from the original API result
+                        # Otherwise, return a new stream that yields the first chunk
+                        # and remaining items
+                        async def combined_stream():  # type: ignore
+                            yield first_chunk
+                            async for chunk in stream_iter:
+                                yield chunk
+
+                        result = combined_stream()  # type: ignore
+                    except StopAsyncIteration:
+                        # Empty stream is normal - nothing to do
+                        pass
+                except Exception as e:
+                    # Any exception here should be raised to trigger the retry mechanism
+                    raise e
+            else:
                 self._cache_store(hashed_key, result.model_dump())
         return cached, hashed_key, result
 

{langroid-0.49.0 → langroid-0.50.0}/langroid/mytypes.py

@@ -65,6 +65,33 @@ class DocMetaData(BaseModel):
 
         return original_dict
 
+    def __str__(self) -> str:
+        title_str = (
+            ""
+            if "unknown" in self.title.lower() or self.title.strip() == ""
+            else f"Title: {self.title}"
+        )
+        date_str = ""
+        if (
+            "unknown" not in self.published_date.lower()
+            and self.published_date.strip() != ""
+        ):
+            try:
+                from dateutil import parser
+
+                # Try to parse the date string
+                date_obj = parser.parse(self.published_date)
+                # Format to include only the date part (year-month-day)
+                date_only = date_obj.strftime("%Y-%m-%d")
+                date_str = f"Date: {date_only}"
+            except (ValueError, ImportError, TypeError):
+                # If parsing fails, just use the original date
+                date_str = f"Date: {self.published_date}"
+        components = [self.source] + (
+            [] if title_str + date_str == "" else [title_str, date_str]
+        )
+        return ", ".join(components)
+
     class Config:
         extra = Extra.allow
 
@@ -93,7 +120,7 @@ class Document(BaseModel):
         return dedent(
             f"""
         CONTENT: {self.content}         
-        SOURCE:{self.metadata.source}
+        SOURCE:{str(self.metadata)}
         """
         )
 

{langroid-0.49.0 → langroid-0.50.0}/langroid/parsing/document_parser.py

@@ -380,9 +380,6 @@ class DocumentParser(Parser):
         Get document chunks from a pdf source,
         with page references in the document metadata.
 
-        Adapted from
-        https://github.com/whitead/paper-qa/blob/main/paperqa/readers.py
-
         Returns:
             List[Document]: a list of `Document` objects,
                 each containing a chunk of text