LangChain and LlamaIndex¶
ReadAnyDoc plugs into both frameworks as a loader. Each page becomes one document by default, as Markdown with tables in HTML, and carries its page number, so answers can point back to the page.
LangChain¶
from readanydoc.integrations.langchain import ReadAnyDocLoader
docs = ReadAnyDocLoader("report.pdf").load() # one Document per page
docs[0].metadata # {"source": "report.pdf", "page": 1}
LlamaIndex¶
from llama_index.core import SimpleDirectoryReader, VectorStoreIndex
from readanydoc.integrations.llamaindex import ReadAnyDocReader
reader = ReadAnyDocReader()
docs = SimpleDirectoryReader("contracts/", file_extractor={".pdf": reader, ".docx": reader}).load_data()
index = VectorStoreIndex.from_documents(docs)
Options¶
| Argument | Values |
|---|---|
split |
"page" (default), "block" (one document per title, paragraph, table...; metadata adds block_type, bbox, confidence), "document" (the whole file) |
any readanydoc.parse argument |
mode="model-only", model=..., pages=[1, 2], ... |
The first document of a file carries warnings in its metadata when there are any (a page the provider declined,
an illegible value). Typed extraction with citations is not part of the loaders; call
readanydoc.extract for that.