langroid 0.1.72__tar.gz → 0.1.76__tar.gz

{langroid-0.1.72 → langroid-0.1.76}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.1.72
+Version: 0.1.76
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani
@@ -70,6 +70,7 @@ Requires-Dist: trafilatura (>=1.5.0,<2.0.0)
 Requires-Dist: typer (>=0.7.0,<0.8.0)
 Requires-Dist: types-redis (>=4.5.5.2,<5.0.0.0)
 Requires-Dist: types-requests (>=2.31.0.1,<3.0.0.0)
+Requires-Dist: unstructured[docx,pdf,pptx] (>=0.10.16,<0.11.0)
 Requires-Dist: wget (>=3.2,<4.0)
 Description-Content-Type: text/markdown
 
@@ -130,8 +131,8 @@ This Multi-Agent paradigm is inspired by the
 [Actor Framework](https://en.wikipedia.org/wiki/Actor_model)
 (but you do not need to know anything about this!). 
 
-Langroid is a fresh take on LLM app-development, where considerable thought has gone 
-into simplifying the developer experience. It does not use `Langchain` or `Llama-Index`.
+`Langroid` is a fresh take on LLM app-development, where considerable thought has gone 
+into simplifying the developer experience; it does not use `Langchain`.
 
 We welcome contributions -- See the [contributions](./CONTRIBUTING.md) document
 for ideas on what to contribute.
@@ -142,6 +143,8 @@ for ideas on what to contribute.
 <summary> <b>:fire: Updates/Releases</b></summary>
 
 - **Sep 2023:**
+  - **0.1.72:** Many improvements to DocChatAgent: better embedding model, 
+          hybrid search to improve retrieval, better pdf parsing, re-ranking retrieved results with cross-encoders. 
   - **Use with local LLama Models:** see tutorial [here](https://langroid.github.io/langroid/blog/2023/09/14/using-langroid-with-local-llms/)
   - **Langroid Blog/Newsletter Launched!**: First post is [here](https://substack.com/notes/post/p-136704592) -- Please subscribe to stay updated. 
   - **0.1.56:** Support Azure OpenAI. 
@@ -167,7 +170,7 @@ See [this test](tests/main/test_recipient_tool.py) for example usage.
   - **0.1.27**: Added [support](langroid/cachedb/momento_cachedb.py) 
     for [Momento Serverless Cache](https://www.gomomento.com/) as an alternative to Redis.
   - **0.1.24**: [`DocChatAgent`](langroid/agent/special/doc_chat_agent.py) 
-    now [accepts](langroid/parsing/pdf_parser.py) PDF files or URLs.
+    now [accepts](langroid/parsing/document_parser.py) PDF files or URLs.
 
 </details>
 
@@ -231,9 +234,6 @@ Here is what it looks like in action:
 
 # :gear: Installation and Setup
 
-:whale: For a simpler setup, see the Docker section below, which lets you get started just
-by setting up environment variables in a `.env` file.
-
 ### Install `langroid`
 Langroid requires Python 3.11+. We recommend using a virtual environment.
 Use `pip` to install `langroid` (from PyPi) to your virtual environment:
@@ -663,6 +663,7 @@ First create a `DocChatAgentConfig` instance, with a
 
 ```python
 from langroid.agent.doc_chat_agent import DocChatAgentConfig
+from langroid.vector_store.qdrantdb import QdrantDBConfig
 config = DocChatAgentConfig(
   doc_paths = [
     "https://en.wikipedia.org/wiki/Language_model",
@@ -672,9 +673,7 @@ config = DocChatAgentConfig(
   llm = OpenAIGPTConfig(
     chat_model=OpenAIChatModel.GPT4,
   ),
-  vecdb=VectorStoreConfig(
-    type="qdrant",
-  ),
+  vecdb=QdrantDBConfig()
 )
 ```
 

{langroid-0.1.72 → langroid-0.1.76}/README.md

@@ -55,8 +55,8 @@ This Multi-Agent paradigm is inspired by the
 [Actor Framework](https://en.wikipedia.org/wiki/Actor_model)
 (but you do not need to know anything about this!). 
 
-Langroid is a fresh take on LLM app-development, where considerable thought has gone 
-into simplifying the developer experience. It does not use `Langchain` or `Llama-Index`.
+`Langroid` is a fresh take on LLM app-development, where considerable thought has gone 
+into simplifying the developer experience; it does not use `Langchain`.
 
 We welcome contributions -- See the [contributions](./CONTRIBUTING.md) document
 for ideas on what to contribute.
@@ -67,6 +67,8 @@ for ideas on what to contribute.
 <summary> <b>:fire: Updates/Releases</b></summary>
 
 - **Sep 2023:**
+  - **0.1.72:** Many improvements to DocChatAgent: better embedding model, 
+          hybrid search to improve retrieval, better pdf parsing, re-ranking retrieved results with cross-encoders. 
   - **Use with local LLama Models:** see tutorial [here](https://langroid.github.io/langroid/blog/2023/09/14/using-langroid-with-local-llms/)
   - **Langroid Blog/Newsletter Launched!**: First post is [here](https://substack.com/notes/post/p-136704592) -- Please subscribe to stay updated. 
   - **0.1.56:** Support Azure OpenAI. 
@@ -92,7 +94,7 @@ See [this test](tests/main/test_recipient_tool.py) for example usage.
   - **0.1.27**: Added [support](langroid/cachedb/momento_cachedb.py) 
     for [Momento Serverless Cache](https://www.gomomento.com/) as an alternative to Redis.
   - **0.1.24**: [`DocChatAgent`](langroid/agent/special/doc_chat_agent.py) 
-    now [accepts](langroid/parsing/pdf_parser.py) PDF files or URLs.
+    now [accepts](langroid/parsing/document_parser.py) PDF files or URLs.
 
 </details>
 
@@ -156,9 +158,6 @@ Here is what it looks like in action:
 
 # :gear: Installation and Setup
 
-:whale: For a simpler setup, see the Docker section below, which lets you get started just
-by setting up environment variables in a `.env` file.
-
 ### Install `langroid`
 Langroid requires Python 3.11+. We recommend using a virtual environment.
 Use `pip` to install `langroid` (from PyPi) to your virtual environment:
@@ -588,6 +587,7 @@ First create a `DocChatAgentConfig` instance, with a
 
 ```python
 from langroid.agent.doc_chat_agent import DocChatAgentConfig
+from langroid.vector_store.qdrantdb import QdrantDBConfig
 config = DocChatAgentConfig(
   doc_paths = [
     "https://en.wikipedia.org/wiki/Language_model",
@@ -597,9 +597,7 @@ config = DocChatAgentConfig(
   llm = OpenAIGPTConfig(
     chat_model=OpenAIChatModel.GPT4,
   ),
-  vecdb=VectorStoreConfig(
-    type="qdrant",
-  ),
+  vecdb=QdrantDBConfig()
 )
 ```
 

{langroid-0.1.72 → langroid-0.1.76}/langroid/agent/special/doc_chat_agent.py

@@ -81,6 +81,7 @@ class DocChatAgentConfig(ChatAgentConfig):
     # and use the embed(A) to find similar chunks in vecdb.
     # Referred to as HyDE in the paper:
     # https://arxiv.org/pdf/2212.10496.pdf
+    # It is False by default; its benefits depends on the context.
     hypothetical_answer: bool = False
     n_query_rephrases: int = 0
     use_fuzzy_match: bool = True
@@ -125,7 +126,6 @@ class DocChatAgentConfig(ChatAgentConfig):
     )
 
     vecdb: VectorStoreConfig = QdrantDBConfig(
-        type="qdrant",
         collection_name=None,
         storage_path=".qdrant/data/",
         embedding=hf_embed_config,
@@ -392,13 +392,14 @@ class DocChatAgent(ChatAgent):
         if self.config.hypothetical_answer:
             with console.status("[cyan]LLM generating hypothetical answer..."):
                 with StreamingIfAllowed(self.llm, False):
+                    # TODO: provide an easy way to
+                    # Adjust this prompt depending on context.
                     answer = self.llm_response_forget(
                         f"""
-                        Give a sample answer to the following query, 
+                        Give an ideal answer to the following query, 
                         in up to 3 sentences. Do not explain yourself, 
                         and do not apologize, just show 
-                        a possible answer. Guess a hypothetical answer 
-                        even if you do not have any information.
+                        a good possible answer, even if you do not have any information.
                         Preface your answer with "HYPOTHETICAL ANSWER: "
                         
                         QUERY: {query}
@@ -505,7 +506,7 @@ class DocChatAgent(ChatAgent):
 
         with console.status("[cyan]LLM Extracting verbatim passages..."):
             with StreamingIfAllowed(self.llm, False):
-                # these are async calls, one per passage
+                # these are async calls, one per passage; turn off streaming
                 extracts = self.llm.get_verbatim_extracts(query, passages)
                 extracts = [e for e in extracts if e.content != NO_ANSWER]
 

{langroid-0.1.72 → langroid-0.1.76}/langroid/agent/special/retriever_agent.py

@@ -41,7 +41,6 @@ class RetrieverAgentConfig(DocChatAgentConfig):
     stream: bool = True  # allow streaming where needed
     max_tokens: int = 10000
     vecdb: VectorStoreConfig = QdrantDBConfig(
-        type="qdrant",
         collection_name=None,
         storage_path=".qdrant/data/",
         embedding=OpenAIEmbeddingsConfig(

langroid-0.1.72/langroid/parsing/pdf_parser.py → langroid-0.1.76/langroid/parsing/document_parser.py

@@ -1,5 +1,6 @@
 import re
 from abc import abstractmethod
+from enum import Enum
 from io import BytesIO
 from typing import Any, Generator, List, Tuple
 
@@ -12,35 +13,56 @@ from langroid.mytypes import DocMetaData, Document
 from langroid.parsing.parser import Parser, ParsingConfig
 
 
-class PdfParser(Parser):
+class DocumentType(str, Enum):
+    PDF = "pdf"
+    DOCX = "docx"
+
+
+class DocumentParser(Parser):
     """
-    Abstract base class for extracting text from PDFs.
+    Abstract base class for extracting text from special types of docs
+    such as PDFs or Docx.
 
     Attributes:
-        source (str): The PDF source, either a URL or a file path.
-        pdf_bytes (BytesIO): BytesIO object containing the PDF data.
+        source (str): The source, either a URL or a file path.
+        doc_bytes (BytesIO): BytesIO object containing the doc data.
     """
 
     @classmethod
-    def create(cls, source: str, config: ParsingConfig) -> "PdfParser":
+    def create(cls, source: str, config: ParsingConfig) -> "DocumentParser":
         """
-        Create a PDF Parser instance based on config.library specified.
+        Create a DocumentParser instance based on source type
+            and config.<source_type>.library specified.
 
         Args:
             source (str): The source of the PDF, either a URL or a file path.
             config (ParserConfig): The parser configuration.
 
         Returns:
-            PdfParser: An instance of a PDF Parser subclass.
-        """
-        if config.pdf.library == "fitz":
-            return FitzPdfParser(source, config)
-        elif config.pdf.library == "pypdf":
-            return PyPdfParser(source, config)
-        elif config.pdf.library == "pdfplumber":
-            return PdfPlumberParser(source, config)
+            DocumentParser: An instance of a DocumentParser subclass.
+        """
+        if DocumentParser._document_type(source) == DocumentType.PDF:
+            if config.pdf.library == "fitz":
+                return FitzPDFParser(source, config)
+            elif config.pdf.library == "pypdf":
+                return PyPDFParser(source, config)
+            elif config.pdf.library == "pdfplumber":
+                return PDFPlumberParser(source, config)
+            elif config.pdf.library == "unstructured":
+                return UnstructuredPDFParser(source, config)
+            else:
+                raise ValueError(
+                    f"Unsupported PDF library specified: {config.pdf.library}"
+                )
+        elif DocumentParser._document_type(source) == DocumentType.DOCX:
+            if config.docx.library == "unstructured":
+                return UnstructuredDocxParser(source, config)
+            else:
+                raise ValueError(
+                    f"Unsupported DOCX library specified: {config.docx.library}"
+                )
         else:
-            raise ValueError(f"Unsupported library specified: {config.pdf.library}")
+            raise ValueError(f"Unsupported document type: {source}")
 
     def __init__(self, source: str, config: ParsingConfig):
         """
@@ -52,14 +74,32 @@ class PdfParser(Parser):
         super().__init__(config)
         self.source = source
         self.config = config
-        self.pdf_bytes = self._load_pdf_as_bytesio()
+        self.doc_bytes = self._load_doc_as_bytesio()
+
+    @staticmethod
+    def _document_type(source: str) -> DocumentType:
+        """
+        Determine the type of document based on the source.
+
+        Args:
+            source (str): The source of the PDF, either a URL or a file path.
+
+        Returns:
+            str: The document type.
+        """
+        if source.lower().endswith(".pdf"):
+            return DocumentType.PDF
+        elif source.lower().endswith(".docx"):
+            return DocumentType.DOCX
+        else:
+            raise ValueError(f"Unsupported document type: {source}")
 
-    def _load_pdf_as_bytesio(self) -> BytesIO:
+    def _load_doc_as_bytesio(self) -> BytesIO:
         """
-        Load the PDF into a BytesIO object.
+        Load the docs into a BytesIO object.
 
         Returns:
-            BytesIO: A BytesIO object containing the PDF data.
+            BytesIO: A BytesIO object containing the doc data.
         """
         if self.source.startswith(("http://", "https://")):
             response = requests.get(self.source)
@@ -159,7 +199,7 @@ class PdfParser(Parser):
         return docs
 
 
-class FitzPdfParser(PdfParser):
+class FitzPDFParser(DocumentParser):
     """
     Parser for processing PDFs using the `fitz` library.
     """
@@ -171,7 +211,7 @@ class FitzPdfParser(PdfParser):
         Returns:
             Generator[fitz.Page]: Generator yielding each page.
         """
-        doc = fitz.open(stream=self.pdf_bytes, filetype="pdf")
+        doc = fitz.open(stream=self.doc_bytes, filetype="pdf")
         for i, page in enumerate(doc):
             yield i, page
         doc.close()
@@ -189,7 +229,7 @@ class FitzPdfParser(PdfParser):
         return self.fix_text(page.get_text())
 
 
-class PyPdfParser(PdfParser):
+class PyPDFParser(DocumentParser):
     """
     Parser for processing PDFs using the `pypdf` library.
     """
@@ -201,7 +241,7 @@ class PyPdfParser(PdfParser):
         Returns:
             Generator[pypdf.pdf.PageObject]: Generator yielding each page.
         """
-        reader = pypdf.PdfReader(self.pdf_bytes)
+        reader = pypdf.PdfReader(self.doc_bytes)
         for i, page in enumerate(reader.pages):
             yield i, page
 
@@ -218,7 +258,7 @@ class PyPdfParser(PdfParser):
         return self.fix_text(page.extract_text())
 
 
-class PdfPlumberParser(PdfParser):
+class PDFPlumberParser(DocumentParser):
     """
     Parser for processing PDFs using the `pdfplumber` library.
     """
@@ -232,7 +272,7 @@ class PdfPlumberParser(PdfParser):
         Returns:
             Generator[pdfplumber.Page]: Generator yielding each page.
         """
-        with pdfplumber.open(self.pdf_bytes) as pdf:
+        with pdfplumber.open(self.doc_bytes) as pdf:
             for i, page in enumerate(pdf.pages):
                 yield i, page
 
@@ -247,3 +287,60 @@ class PdfPlumberParser(PdfParser):
             str: Extracted text from the page.
         """
         return self.fix_text(page.extract_text())
+
+
+class UnstructuredPDFParser(DocumentParser):
+    """
+    Parser for processing PDF files using the `unstructured` library.
+    """
+
+    def iterate_pages(self) -> Generator[Tuple[int, Any], None, None]:  # type: ignore
+        from unstructured.partition.pdf import partition_pdf
+
+        elements = partition_pdf(file=self.doc_bytes, include_page_breaks=True)
+        for i, el in enumerate(elements):
+            yield i, el
+
+    def extract_text_from_page(self, page: Any) -> str:
+        """
+        Extract text from a given `unstructured` element.
+
+        Args:
+            page (unstructured element): The `unstructured` element object.
+
+        Returns:
+            str: Extracted text from the element.
+        """
+        return self.fix_text(str(page))
+
+
+class UnstructuredDocxParser(DocumentParser):
+    """
+    Parser for processing DOCX files using the `unstructured` library.
+    """
+
+    def iterate_pages(self) -> Generator[Tuple[int, Any], None, None]:  # type: ignore
+        from unstructured.partition.docx import partition_docx
+
+        elements = partition_docx(file=self.doc_bytes)
+        for i, el in enumerate(elements):
+            yield i, el
+
+    def extract_text_from_page(self, page: Any) -> str:
+        """
+        Extract text from a given `unstructured` element.
+
+        Note:
+            The concept of "pages" doesn't actually exist in the .docx file format in
+            the same way it does in formats like .pdf. A .docx file is made up of a
+            series of elements like paragraphs and tables, but the division into
+            pages is done dynamically based on the rendering settings (like the page
+            size, margin size, font size, etc.).
+
+        Args:
+            page (unstructured element): The `unstructured` element object.
+
+        Returns:
+            str: Extracted text from the element.
+        """
+        return self.fix_text(str(page))

{langroid-0.1.72 → langroid-0.1.76}/langroid/parsing/parser.py

@@ -23,6 +23,10 @@ class PdfParsingConfig(BaseSettings):
     library: str = "pdfplumber"
 
 
+class DocxParsingConfig(BaseSettings):
+    library: str = "unstructured"
+
+
 class ParsingConfig(BaseSettings):
     splitter: str = Splitter.TOKENS
     chunk_size: int = 200  # aim for this many tokens per chunk
@@ -35,6 +39,7 @@ class ParsingConfig(BaseSettings):
     separators: List[str] = ["\n\n", "\n", " ", ""]
     token_encoding_model: str = "text-embedding-ada-002"
     pdf: PdfParsingConfig = PdfParsingConfig()
+    docx: DocxParsingConfig = DocxParsingConfig()
 
 
 class Parser:

{langroid-0.1.72 → langroid-0.1.76}/langroid/parsing/repo_loader.py

@@ -18,8 +18,8 @@ from github.Repository import Repository
 from pydantic import BaseSettings
 
 from langroid.mytypes import DocMetaData, Document
-from langroid.parsing.parser import Parser
-from langroid.parsing.pdf_parser import PdfParser
+from langroid.parsing.document_parser import DocumentParser
+from langroid.parsing.parser import Parser, ParsingConfig
 
 logger = logging.getLogger(__name__)
 
@@ -440,7 +440,7 @@ class RepoLoader:
     @staticmethod
     def get_documents(
         path: str,
-        parser: Parser,
+        parser: Parser = Parser(ParsingConfig()),
         file_types: Optional[List[str]] = None,
         exclude_dirs: Optional[List[str]] = None,
         depth: int = -1,
@@ -493,12 +493,12 @@ class RepoLoader:
 
         for file_path in file_paths:
             _, file_extension = os.path.splitext(file_path)
-            if file_extension.lower() == ".pdf":
-                pdf_parser = PdfParser.create(
+            if file_extension.lower() in [".pdf", ".docx"]:
+                doc_parser = DocumentParser.create(
                     file_path,
                     parser.config,
                 )
-                docs.extend(pdf_parser.get_doc_chunks())
+                docs.extend(doc_parser.get_doc_chunks())
             else:
                 with open(file_path, "r") as f:
                     if lines is not None:

{langroid-0.1.72 → langroid-0.1.76}/langroid/parsing/url_loader.py

@@ -9,8 +9,8 @@ from trafilatura.downloads import (
 )
 
 from langroid.mytypes import DocMetaData, Document
-from langroid.parsing.parser import Parser
-from langroid.parsing.pdf_parser import PdfParser
+from langroid.parsing.document_parser import DocumentParser
+from langroid.parsing.parser import Parser, ParsingConfig
 
 logging.getLogger("trafilatura").setLevel(logging.ERROR)
 
@@ -27,7 +27,7 @@ class URLLoader:
      the "accept" button on the cookie dialog.
     """
 
-    def __init__(self, urls: List[str], parser: Parser):
+    def __init__(self, urls: List[str], parser: Parser = Parser(ParsingConfig())):
         self.urls = urls
         self.parser = parser
 
@@ -44,12 +44,12 @@ class URLLoader:
                 sleep_time=5,
             )
             for url, result in buffered_downloads(buffer, threads):
-                if url.lower().endswith(".pdf"):
-                    pdf_parser = PdfParser.create(
+                if url.lower().endswith(".pdf") or url.lower().endswith(".docx"):
+                    doc_parser = DocumentParser.create(
                         url,
                         self.parser.config,
                     )
-                    docs.extend(pdf_parser.get_doc_chunks())
+                    docs.extend(doc_parser.get_doc_chunks())
                 else:
                     text = trafilatura.extract(
                         result,

{langroid-0.1.72 → langroid-0.1.76}/langroid/vector_store/base.py

@@ -14,6 +14,7 @@ logger = logging.getLogger(__name__)
 
 
 class VectorStoreConfig(BaseSettings):
+    type: str = "qdrant"  # deprecated, keeping it for backward compatibility
     collection_name: str | None = None
     replace_collection: bool = False  # replace collection if it already exists
     storage_path: str = ".qdrant/data"
@@ -23,7 +24,6 @@ class VectorStoreConfig(BaseSettings):
         model_type="openai",
     )
     timeout: int = 60
-    type: str = "qdrant"
     host: str = "127.0.0.1"
     port: int = 6333
     # compose_file: str = "langroid/vector_store/docker-compose-qdrant.yml"
@@ -38,7 +38,7 @@ class VectorStore(ABC):
         self.config = config
 
     @staticmethod
-    def create(config: VectorStoreConfig) -> "VectorStore":
+    def create(config: VectorStoreConfig) -> Optional["VectorStore"]:
         from langroid.vector_store.chromadb import ChromaDB, ChromaDBConfig
         from langroid.vector_store.qdrantdb import QdrantDB, QdrantDBConfig
 
@@ -47,7 +47,16 @@ class VectorStore(ABC):
         elif isinstance(config, ChromaDBConfig):
             return ChromaDB(config)
         else:
-            raise ValueError(f"Unknown vector store config: {config.__repr_name__()}")
+            logger.warning(
+                f"""
+                Unknown vector store config: {config.__repr_name__()},
+                so skipping vector store creation!
+                If you intended to use a vector-store, please set a specific 
+                vector-store in your script, typically in the `vecdb` field of a 
+                `ChatAgentConfig`, otherwise set it to None.
+                """
+            )
+            return None
 
     @abstractmethod
     def clear_empty_collections(self) -> int:
@@ -123,4 +132,4 @@ class VectorStore(ABC):
     def show_if_debug(self, doc_score_pairs: List[Tuple[Document, float]]) -> None:
         if settings.debug:
             for i, (d, s) in enumerate(doc_score_pairs):
-                print_long_text("red", "italic red", f"MATCH-{i}", d.content)
+                print_long_text("red", "italic red", f"\nMATCH-{i}\n", d.content)

{langroid-0.1.72 → langroid-0.1.76}/langroid/vector_store/chromadb.py

@@ -17,7 +17,6 @@ logger = logging.getLogger(__name__)
 
 
 class ChromaDBConfig(VectorStoreConfig):
-    type: str = "chroma"
     collection_name: str = "chroma-langroid"
     storage_path: str = ".chroma/data"
     embedding: EmbeddingModelsConfig = OpenAIEmbeddingsConfig()

{langroid-0.1.72 → langroid-0.1.76}/langroid/vector_store/qdrantdb.py

@@ -28,9 +28,7 @@ logger = logging.getLogger(__name__)
 
 
 class QdrantDBConfig(VectorStoreConfig):
-    type: str = "qdrant"
     cloud: bool = True
-
     collection_name: str | None = None
     storage_path: str = ".qdrant/data"
     embedding: EmbeddingModelsConfig = OpenAIEmbeddingsConfig()

{langroid-0.1.72 → langroid-0.1.76}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.1.72"
+version = "0.1.76"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"
@@ -54,6 +54,7 @@ sqlalchemy = "^2.0.19"
 prettytable = "^3.8.0"
 google-api-python-client = "^2.95.0"
 lxml = "^4.9.3"
+unstructured = {extras = ["docx", "pptx", "pdf"], version = "^0.10.16"}
 
 # optional; see extras section below
 sentence-transformers = {version="2.2.2", optional=true}