mirror of
https://github.com/DS4SD/docling.git
synced 2025-07-26 20:14:47 +00:00
506 lines
14 KiB
Plaintext
506 lines
14 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"# RAG with Docling and 🦜🔗 LangChain"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Note: you may need to restart the kernel to use updated packages.\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"# requirements for this example:\n",
|
|
"%pip install -qq docling docling-core python-dotenv langchain langchain-text-splitters langchain-huggingface langchain-milvus"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"data": {
|
|
"text/plain": [
|
|
"True"
|
|
]
|
|
},
|
|
"execution_count": 2,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": [
|
|
"import os\n",
|
|
"\n",
|
|
"from dotenv import load_dotenv\n",
|
|
"\n",
|
|
"load_dotenv()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 3,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import warnings\n",
|
|
"\n",
|
|
"warnings.filterwarnings(action=\"ignore\", category=UserWarning, module=\"pydantic|torch\")\n",
|
|
"warnings.filterwarnings(action=\"ignore\", category=FutureWarning, module=\"easyocr\")\n",
|
|
"# https://github.com/huggingface/transformers/issues/5486:\n",
|
|
"os.environ[\"TOKENIZERS_PARALLELISM\"] = \"false\""
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Setup"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Helpers"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"Below we set up:\n",
|
|
"- a `Loader` which will be used to create LangChain documents,\n",
|
|
"- a splitter, which will be used to split these documents, and\n",
|
|
"- a helper function for QA printing"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 4,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"from enum import Enum\n",
|
|
"from typing import Iterator\n",
|
|
"\n",
|
|
"from langchain_core.document_loaders import BaseLoader\n",
|
|
"from langchain_core.documents import Document as LCDocument\n",
|
|
"\n",
|
|
"from docling.document_converter import DocumentConverter\n",
|
|
"\n",
|
|
"_KEY_DL_DOC_HASH = \"dl_doc_hash\"\n",
|
|
"_KEY_ORIGIN = \"origin\"\n",
|
|
"\n",
|
|
"\n",
|
|
"class DoclingPDFLoader(BaseLoader):\n",
|
|
" class ParseType(str, Enum):\n",
|
|
" MARKDOWN = \"markdown\"\n",
|
|
" JSON = \"json\"\n",
|
|
"\n",
|
|
" include_origin: bool = False\n",
|
|
"\n",
|
|
" def __init__(self, file_path: str | list[str], parse_type: ParseType) -> None:\n",
|
|
" self._file_paths = file_path if isinstance(file_path, list) else [file_path]\n",
|
|
" self._parse_type = parse_type\n",
|
|
" self._converter = DocumentConverter()\n",
|
|
"\n",
|
|
" def lazy_load(self) -> Iterator[LCDocument]:\n",
|
|
" for source in self._file_paths:\n",
|
|
" dl_doc = self._converter.convert_single(source).output\n",
|
|
" match self._parse_type:\n",
|
|
" case self.ParseType.MARKDOWN:\n",
|
|
" text = dl_doc.export_to_markdown()\n",
|
|
" case self.ParseType.JSON:\n",
|
|
" text = dl_doc.model_dump_json()\n",
|
|
" case _:\n",
|
|
" raise RuntimeError(\n",
|
|
" f\"Unexpected parse type encountered: {self._parse_type}\"\n",
|
|
" )\n",
|
|
" metadata = {\n",
|
|
" _KEY_DL_DOC_HASH: dl_doc.file_info.document_hash,\n",
|
|
" }\n",
|
|
" if self.include_origin:\n",
|
|
" metadata[_KEY_ORIGIN] = source\n",
|
|
"\n",
|
|
" lc_doc = LCDocument(\n",
|
|
" page_content=text,\n",
|
|
" metadata=metadata,\n",
|
|
" )\n",
|
|
" yield lc_doc"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 5,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import json\n",
|
|
"from typing import Iterable, List\n",
|
|
"\n",
|
|
"from docling_core.transforms.chunker import BaseChunker, HierarchicalChunker\n",
|
|
"from docling_core.types import Document as DLDocument\n",
|
|
"from langchain_core.documents import Document as LCDocument\n",
|
|
"\n",
|
|
"\n",
|
|
"class DoclingSplitter:\n",
|
|
"\n",
|
|
" def __init__(\n",
|
|
" self,\n",
|
|
" chunker: BaseChunker | None = None,\n",
|
|
" ) -> None:\n",
|
|
" self.chunker: BaseChunker = chunker or HierarchicalChunker(\n",
|
|
" heading_as_metadata=True\n",
|
|
" )\n",
|
|
"\n",
|
|
" def split_documents(self, documents: Iterable[LCDocument]) -> List[LCDocument]:\n",
|
|
"\n",
|
|
" all_chunk_docs: list[LCDocument] = []\n",
|
|
" for doc in documents:\n",
|
|
" lc_doc: LCDocument = LCDocument.parse_obj(doc)\n",
|
|
" dl_doc: DLDocument = DLDocument.model_validate_json(lc_doc.page_content)\n",
|
|
" chunk_iter = self.chunker.chunk(dl_doc=dl_doc)\n",
|
|
" for chunk in chunk_iter:\n",
|
|
" chunk_metadata = chunk.model_dump(\n",
|
|
" exclude=\"text\",\n",
|
|
" exclude_none=True,\n",
|
|
" )\n",
|
|
" metadata = {**lc_doc.metadata, **chunk_metadata}\n",
|
|
" for k, v in metadata.items():\n",
|
|
" if isinstance(v, Iterable) and not isinstance(v, str):\n",
|
|
" metadata[k] = json.dumps(v)\n",
|
|
" chunk_doc = LCDocument(\n",
|
|
" page_content=chunk.text,\n",
|
|
" metadata=metadata,\n",
|
|
" )\n",
|
|
" all_chunk_docs.append(chunk_doc)\n",
|
|
"\n",
|
|
" return all_chunk_docs"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 6,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"def print_qa(resp_dict):\n",
|
|
" def clip(inp, max_len=100):\n",
|
|
" if isinstance(inp, str):\n",
|
|
" return f\"{inp[:max_len]}{'...' if len(inp) > max_len else ''}\"\n",
|
|
" else:\n",
|
|
" return inp\n",
|
|
"\n",
|
|
" print(\n",
|
|
" f\"Question:\\n{resp_dict['input']}\\n\\nAnswer:\\n{json.dumps(clip(resp_dict['answer']))}\"\n",
|
|
" )\n",
|
|
" for i, doc in enumerate(resp_dict[\"context\"]):\n",
|
|
" print()\n",
|
|
" print(f\"Source {i+1}:\")\n",
|
|
" print(f\" text: {json.dumps(clip(doc.page_content))}\")\n",
|
|
" for key in doc.metadata:\n",
|
|
" if key != \"pk\":\n",
|
|
" print(f\" {key}: {clip(doc.metadata.get(key))}\")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 7,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"FILE_PATH = \"https://arxiv.org/pdf/2206.01062\" # DocLayNet paper"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Loader and splitter"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"**Using native Docling format (as JSON)**"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"To leverage Docling's rich document structure format, we can namely export to JSON and use the `DoclingSplitter` accordingly:"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 8,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"loader = DoclingPDFLoader(\n",
|
|
" file_path=FILE_PATH,\n",
|
|
" parse_type=DoclingPDFLoader.ParseType.JSON,\n",
|
|
")\n",
|
|
"splitter = DoclingSplitter()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"**Using Markdown:**"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"Alternatively, to just use the flat Markdown export instead of the native document format, one can uncomment and use the following:"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 9,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"# from langchain_text_splitters import RecursiveCharacterTextSplitter\n",
|
|
"\n",
|
|
"# loader = DoclingPDFLoader(\n",
|
|
"# file_path=FILE_PATH,\n",
|
|
"# parse_type=DoclingPDFLoader.ParseType.MARKDOWN,\n",
|
|
"# )\n",
|
|
"# splitter = RecursiveCharacterTextSplitter(\n",
|
|
"# chunk_size=1000,\n",
|
|
"# chunk_overlap=200,\n",
|
|
"# )"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"We now used the above-defined objects to get the document splits:"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 10,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"docs = loader.load()\n",
|
|
"splits = splitter.split_documents(docs)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Embed model"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 11,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"from langchain_huggingface.embeddings import HuggingFaceEmbeddings\n",
|
|
"\n",
|
|
"HF_EMBED_MODEL_ID = \"BAAI/bge-small-en-v1.5\"\n",
|
|
"embedding = HuggingFaceEmbeddings(model_name=HF_EMBED_MODEL_ID)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Vector store"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 12,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"from tempfile import TemporaryDirectory\n",
|
|
"\n",
|
|
"from langchain_milvus import Milvus\n",
|
|
"\n",
|
|
"MILVUS_URI = os.environ.get(\n",
|
|
" \"MILVUS_URI\", f\"{(tmp_dir := TemporaryDirectory()).name}/milvus_demo.db\"\n",
|
|
")\n",
|
|
"\n",
|
|
"vectorstore = Milvus.from_documents(\n",
|
|
" splits,\n",
|
|
" embedding,\n",
|
|
" connection_args={\"uri\": MILVUS_URI},\n",
|
|
" collection_name=\"docling_lc_demo\",\n",
|
|
" drop_old=True,\n",
|
|
")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### LLM"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 13,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"The token has not been saved to the git credentials helper. Pass `add_to_git_credential=True` in this function directly or `--add-to-git-credential` if using via `huggingface-cli` if you want to set the git credential as well.\n",
|
|
"Token is valid (permission: write).\n",
|
|
"Your token has been saved to /Users/pva/.cache/huggingface/token\n",
|
|
"Login successful\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"from langchain_huggingface import HuggingFaceEndpoint\n",
|
|
"\n",
|
|
"HF_API_KEY = os.environ.get(\"HF_API_KEY\")\n",
|
|
"HF_LLM_MODEL_ID = \"mistralai/Mistral-7B-Instruct-v0.3\"\n",
|
|
"\n",
|
|
"llm = HuggingFaceEndpoint(\n",
|
|
" repo_id=HF_LLM_MODEL_ID,\n",
|
|
" huggingfacehub_api_token=HF_API_KEY,\n",
|
|
")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## RAG"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 14,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"from langchain.chains import create_retrieval_chain\n",
|
|
"from langchain.chains.combine_documents import create_stuff_documents_chain\n",
|
|
"from langchain_core.prompts import PromptTemplate\n",
|
|
"\n",
|
|
"retriever = vectorstore.as_retriever()\n",
|
|
"prompt = PromptTemplate.from_template(\n",
|
|
" \"Context information is below.\\n---------------------\\n{context}\\n---------------------\\nGiven the context information and not prior knowledge, answer the query.\\nQuery: {input}\\nAnswer:\\n\"\n",
|
|
")\n",
|
|
"question_answer_chain = create_stuff_documents_chain(llm, prompt)\n",
|
|
"rag_chain = create_retrieval_chain(retriever, question_answer_chain)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 15,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Question:\n",
|
|
"How many pages were human annotated by humans for DocLayNet?\n",
|
|
"\n",
|
|
"Answer:\n",
|
|
"\"80863 pages were annotated by humans in DocLayNet.\"\n",
|
|
"\n",
|
|
"Source 1:\n",
|
|
" text: \"DocLayNet contains 80863 PDF pages. Among these, 7059 carry two instances of human annotations, and ...\"\n",
|
|
" bbox: [317.2852478027344, 116.46983337402344, 559.7131958007812, 201.73675537109375]\n",
|
|
" dl_doc_hash: 5dfbd8c115a15fd3396b68409124cfee29fc8efac7b5c846634ff924e635e0dc\n",
|
|
" heading: 3 THE DOCLAYNET DATASET\n",
|
|
" page: 2\n",
|
|
" path: $.main-text[37]\n",
|
|
"\n",
|
|
"Source 2:\n",
|
|
" text: \"In this paper, we present the DocLayNet dataset. It provides pageby-page layout annotation ground-tr...\"\n",
|
|
" bbox: [53.50020980834961, 212.36782836914062, 295.56396484375, 286.4964599609375]\n",
|
|
" dl_doc_hash: 5dfbd8c115a15fd3396b68409124cfee29fc8efac7b5c846634ff924e635e0dc\n",
|
|
" heading: 1 INTRODUCTION\n",
|
|
" page: 2\n",
|
|
" path: $.main-text[23]\n",
|
|
"\n",
|
|
"Source 3:\n",
|
|
" text: \"DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis\"\n",
|
|
" bbox: [53.60108947753906, 723.3781127929688, 347.139892578125, 731.6909790039062]\n",
|
|
" dl_doc_hash: 5dfbd8c115a15fd3396b68409124cfee29fc8efac7b5c846634ff924e635e0dc\n",
|
|
" heading: REFERENCES\n",
|
|
" page: 9\n",
|
|
" path: $.main-text[133]\n",
|
|
"\n",
|
|
"Source 4:\n",
|
|
" text: \"DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis\"\n",
|
|
" bbox: [53.542964935302734, 723.3500366210938, 347.0172424316406, 731.6931762695312]\n",
|
|
" dl_doc_hash: 5dfbd8c115a15fd3396b68409124cfee29fc8efac7b5c846634ff924e635e0dc\n",
|
|
" heading: 4 ANNOTATION CAMPAIGN\n",
|
|
" page: 5\n",
|
|
" path: $.main-text[64]\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"resp_dict = rag_chain.invoke(\n",
|
|
" {\"input\": \"How many pages were human annotated by humans for DocLayNet?\"}\n",
|
|
")\n",
|
|
"print_qa(resp_dict=resp_dict)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": []
|
|
}
|
|
],
|
|
"metadata": {
|
|
"kernelspec": {
|
|
"display_name": ".venv",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.12.4"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 2
|
|
}
|