Custom document loaders in LangChain

Introduction

Custom document loaders help you bring in data from places not covered by built-in loaders. They let you read and prepare your own files or sources for your app.

You have files in a special format that no default loader supports.

You want to load documents from a private database or API.

You need to preprocess or clean data before using it.

You want to combine multiple sources into one loader.

You want to add custom metadata while loading documents.

Syntax

LangChain

from langchain.document_loaders import BaseLoader

class MyLoader(BaseLoader):
    def __init__(self, source_path: str):
        self.source_path = source_path

    def load(self):
        # read and process your data here
        documents = []
        with open(self.source_path, 'r', encoding='utf-8') as f:
            text = f.read()
            # create Document objects or dicts
            documents.append({'page_content': text, 'metadata': {}})
        return documents

Custom loaders must inherit from BaseLoader and implement a load method.

The load method returns a list of documents, each with content and optional metadata.

Examples

Loads a plain text file and returns its content as one document.

LangChain

class TxtLoader(BaseLoader):
    def __init__(self, filepath):
        self.filepath = filepath

    def load(self):
        with open(self.filepath, 'r', encoding='utf-8') as f:
            text = f.read()
        return [{'page_content': text, 'metadata': {}}]

Loads JSON data and creates documents from each item with metadata.

LangChain

class JsonLoader(BaseLoader):
    def __init__(self, filepath):
        self.filepath = filepath

    def load(self):
        import json
        with open(self.filepath, 'r', encoding='utf-8') as f:
            data = json.load(f)
        docs = []
        for item in data['items']:
            docs.append({'page_content': item['text'], 'metadata': {'id': item['id']}})
        return docs

Sample Program

This example shows a simple custom loader that reads a text file and adds the file path as metadata. It prints the content and metadata of the loaded document.

LangChain

from langchain.document_loaders import BaseLoader

class SimpleTxtLoader(BaseLoader):
    def __init__(self, filepath):
        self.filepath = filepath

    def load(self):
        with open(self.filepath, 'r', encoding='utf-8') as f:
            text = f.read()
        return [{'page_content': text, 'metadata': {'source': self.filepath}}]

# Usage example
loader = SimpleTxtLoader('example.txt')
docs = loader.load()
for doc in docs:
    print(f"Content:\n{doc['page_content']}")
    print(f"Metadata: {doc['metadata']}")

Important Notes

Make sure your custom loader handles file encoding and errors gracefully.

Adding metadata helps track where documents come from and can be useful later.

Test your loader with different inputs to ensure it works as expected.

Summary

Custom document loaders let you bring in data from any source you want.

They must inherit from BaseLoader and implement a load method.

Use them to read, clean, and add metadata to your documents before using them.

Practice

1. What is the main purpose of creating a custom document loader in Langchain?

easy

A. To load and preprocess documents from any data source before use

B. To create user interfaces for document display

C. To train machine learning models directly

D. To manage database connections automatically

Solution

1. Step 1: Understand the role of document loaders

   Document loaders bring data into Langchain, preparing it for processing.

2. Step 2: Identify the purpose of custom loaders

   Custom loaders allow loading from any source and preprocessing like cleaning or adding metadata.

3. Final Answer:

   To load and preprocess documents from any data source before use -> Option A

4. Quick Check:

   Custom loaders = load + preprocess documents [OK]

Hint: Custom loaders bring in and prep data from any source [OK]

Common Mistakes:

• Confusing loaders with UI components
• Thinking loaders train models
• Assuming loaders manage databases

2. Which of the following is the correct way to start defining a custom document loader class in Langchain?

easy

A. class MyLoader():\n def load(self):\n pass

B. def MyLoader():\n return BaseLoader.load()

C. class MyLoader(BaseLoader):\n def load(self):\n pass

D. class MyLoader(BaseLoader):\n def read(self):\n pass

Solution

1. Step 1: Check inheritance from BaseLoader

   Custom loaders must inherit from BaseLoader to fit Langchain's pattern.

2. Step 2: Confirm the method name is load

   The required method to implement is load(), not read().

3. Final Answer:

   class MyLoader(BaseLoader):\n def load(self):\n pass -> Option C

4. Quick Check:

   Inherit BaseLoader + define load() method [OK]

Hint: Inherit BaseLoader and implement load() method [OK]

Common Mistakes:

• Not inheriting from BaseLoader
• Using wrong method name like read()
• Defining function instead of class

3. Given this custom loader code snippet, what will print(docs) output?

from langchain.schema import Document
from langchain.document_loaders.base import BaseLoader

class SimpleLoader(BaseLoader):
    def load(self):
        return [Document(page_content='Hello World', metadata={'source': 'test'})]

loader = SimpleLoader()
docs = loader.load()

medium

A. [Document(page_content='Hello World', metadata={'source': 'test'})]

B. ['Hello World']

C. None

D. SyntaxError

Solution

1. Step 1: Analyze the load method return value

   The load method returns a list with one Document object containing content and metadata.

2. Step 2: Understand what docs holds

   docs is assigned the list returned, so printing docs shows the list with the Document instance.

3. Final Answer:

   [Document(page_content='Hello World', metadata={'source': 'test'})] -> Option A

4. Quick Check:

   load() returns list of Document objects [OK]

Hint: load() returns list of Document objects, not strings [OK]

Common Mistakes:

• Expecting a list of strings instead of Document objects
• Thinking load() returns None
• Assuming syntax error due to imports

4. Identify the error in this custom loader code:

from langchain.document_loaders.base import BaseLoader

class FaultyLoader(BaseLoader):
    def load(self):
        docs = []
        docs.append('Just a string')
        return docs

loader = FaultyLoader()
loader.load()

medium

A. The load method must be named read

B. BaseLoader cannot be inherited directly

C. The class must not have a load method

D. The load method should return Document objects, not strings

Solution

1. Step 1: Check what load returns

   The load method returns a list with a string instead of Document objects.

2. Step 2: Recall Langchain expects Document objects

   Custom loaders must return Document instances, not plain strings, for proper processing.

3. Final Answer:

   The load method should return Document objects, not strings -> Option D

4. Quick Check:

   load() must return Document objects [OK]

Hint: load() must return Document objects, not plain strings [OK]

Common Mistakes:

• Returning strings instead of Document objects
• Misnaming the load method
• Thinking BaseLoader can't be inherited

5. You want to create a custom loader that reads text files from a folder, adds the filename as metadata, and returns Document objects. Which approach correctly implements this?

hard

A. Use BaseLoader directly without subclassing, call load() with folder path argument

B. Inherit BaseLoader, implement load() to read files, create Document with content and metadata filename, return list

C. Create a function that returns file contents as strings without Document wrapping

D. Subclass BaseLoader but return raw file paths as strings in load()

Solution

1. Step 1: Confirm inheritance and method implementation

   Custom loaders must subclass BaseLoader and implement load() method.

2. Step 2: Ensure load reads files and returns Document objects with metadata

   Reading files and wrapping content in Document with filename metadata is correct.

3. Final Answer:

   Inherit BaseLoader, implement load() to read files, create Document with content and metadata filename, return list -> Option B

4. Quick Check:

   Subclass + load() returns Document list with metadata [OK]

Hint: Subclass BaseLoader, load files, wrap in Document with metadata [OK]

Common Mistakes:

• Not subclassing BaseLoader
• Returning strings instead of Document objects
• Ignoring metadata addition