Skip to main content


Nuclia automatically indexes your unstructured data from any internal and external source, providing optimized search results and generative answers. It can handle video and audio transcription, image content extraction, and document parsing.

The Nuclia Understanding API supports the processing of unstructured data, including text, web pages, documents, and audio/video contents. It extracts all texts wherever they are (using speech-to-text or OCR when needed), it also extracts metadata, embedded files (like images in a PDF), and web links. If machine learning is enabled, it identifies entities, provides a summary of the content and generates embeddings for all the sentences.


To use the Nuclia Understanding API, you need to have a Nuclia account. You can create one for free at, and then create a NUA key.

#!pip install --upgrade protobuf
#!pip install nucliadb-protos
import os

os.environ["NUCLIA_ZONE"] = "<YOUR_ZONE>" # e.g. europe-1
os.environ["NUCLIA_NUA_KEY"] = "<YOUR_API_KEY>"


To use the Nuclia document loader, you need to instantiate a NucliaUnderstandingAPI tool:

from import NucliaUnderstandingAPI

nua = NucliaUnderstandingAPI(enable_ml=False)
from langchain.document_loaders.nuclia import NucliaLoader

loader = NucliaLoader("./interview.mp4", nua)

You can now call the load the document in a loop until you get the document.

import time

pending = True
while pending:
docs = loader.load()
if len(docs) > 0:
pending = False

Retrieved information

Nuclia returns the following information:

  • file metadata
  • extracted text
  • nested text (like text in an embedded image)
  • paragraphs and sentences splitting (defined by the position of their first and last characters, plus start time and end time for a video or audio file)
  • links
  • a thumbnail
  • embedded files


Generated files (thumbnail, extracted embedded files, etc.) are provided as a token. You can download them with the /processing/download endpoint.

Also at any level, if an attribute exceeds a certain size, it will be put in a downloadable file and will be replaced in the document by a file pointer. This will consist of {"file": {"uri": "JWT_TOKEN"}}. The rule is that if the size of the message is greater than 1000000 characters, the biggest parts will be moved to downloadable files. First, the compression process will target vectors. If that is not enough, it will target large field metadata, and finally it will target extracted text.