LangExtract + Milvus Integration

This guide demonstrates how to use LangExtract with Milvus to build an intelligent document processing and retrieval system.

LangExtract is a Python library that uses Large Language Models (LLMs) to extract structured information from unstructured text documents with precise source grounding. The system combines LangExtract's extraction capabilities with Milvus's vector storage to enable both semantic similarity search and precise metadata filtering.

This integration is particularly valuable for content management, semantic search, knowledge discovery, and building recommendation systems based on extracted document attributes.

Prerequisites

Before running this notebook, make sure you have the following dependencies installed:

If you are using Google Colab, to enable dependencies just installed, you may need to restart the runtime (click on the "Runtime" menu at the top of the screen, and select "Restart session" from the dropdown menu).

We will use Gemini as the LLM in this example. You should prepare the api key GEMINI_API_KEY as an environment variable.

Define the LangExtract + Milvus pipeline

We will define the pipeline that uses LangExtract for structured information extraction and Milvus as the vector store.

Configuration and Setup

Let's configure our global parameters for the integration. We'll use Gemini's embedding model to generate vector representations for our documents.

Initialize Milvus Client

Now let's initialize our Milvus client. We'll use a local database file for simplicity, but this can easily be scaled to a full Milvus server deployment.

As for the argument of MilvusClient:

• Setting the uri as a local file, e.g../milvus.db, is the most convenient method, as it automatically utilizes Milvus Lite to store all data in this file.
• If you have large scale of data, you can set up a more performant Milvus server on docker or kubernetes. In this setup, please use the server uri, e.g.http://localhost:19530, as your uri.
• If you want to use Zilliz Cloud, the fully managed cloud service for Milvus, adjust the uri and token, which correspond to the Public Endpoint and Api key in Zilliz Cloud.

Sample Data Preparation

For this demonstration, we'll use movie descriptions as our sample documents. This showcases LangExtract's ability to extract structured information like genres, characters, and themes from unstructured text.

Setting Up the Milvus Collection

Before we can store our extracted data, we need to create a Milvus collection with the appropriate schema. This collection will store the original document text, vector embeddings, and extracted metadata fields.

Defining the Extraction Schema

LangExtract uses prompts and examples to guide the LLM in extracting structured information. Let's define our extraction schema for movie descriptions, specifying what information to extract and how to categorize it.

Providing Examples for Better Extraction

To improve the quality and consistency of extractions, we'll provide LangExtract with a few examples. These examples demonstrate the expected format and help the model understand our extraction requirements.

Processing and Vectorizing the Results

Now we need to process the extraction results and generate vector embeddings for each document. We'll also flatten the extracted attributes into separate fields to make them easily searchable in Milvus.

Inserting Data into Milvus

With our processed data ready, let's insert it into the Milvus collection. This will enable us to perform both semantic searches and precise metadata filtering.

Demonstrating Metadata Filtering

One of the key advantages of combining LangExtract with Milvus is the ability to perform precise filtering based on extracted metadata. Let's demonstrate this with some filter expression searches.

Combining Semantic Search with Metadata Filtering

The real power of this integration comes from combining semantic vector search with precise metadata filtering. This allows us to find semantically similar content while applying specific constraints based on extracted attributes.