Get feedback on Indian addresses

The challenge: Understanding Indian address quality

India's addressing system is known for its diversity and complexity. Addresses can be highly descriptive, non-standardized, include local landmarks, and often lack precise PIN codes or a consistent component order. This poses significant challenges for individuals, ecommerce platforms, logistics companies, and service providers who rely on accurate location data. Key issues often encountered with Indian addresses include:

  • Missing or incorrect Pin codes: Essential for efficient mail and package routing, yet frequently inaccurate or absent.
  • Spelling errors: Common mistakes in the names of localities, cities, or states can lead to misinterpretation.
  • Non-standard component order: The sequence of address elements (like house number, street, locality, city) can vary widely, making automated processing difficult.
  • Lack of standardization: Colloquial terms, abbreviations, and descriptive references (e.g., "near the old temple") are common but not understood by standard systems.
  • Inclusion of relational information: Terms like "S/o" (Son of), "D/o" (Daughter of), or "C/o" (Care of) are frequently embedded within the address, adding non-locational data.
  • Variations in sub-premise notations: Components like unit numbers, house numbers, or plot numbers (e.g., "2/1") are written in numerous ways, such as "2/1", "2-1", "2 by 1", or "No 2, 1st part", making them hard to parse consistently.
  • Prevalence of sub-premise details: Many addresses, especially in urban areas, include crucial sub-premise information like apartment numbers, flat numbers, or building wing details, which are vital for last-mile delivery but often unstructured.
  • Ambiguity: Addresses can sometimes be interpreted in multiple ways, leading to uncertainty in locating the exact point.

These challenges can result in delivery failures, increased operational costs, poor customer experiences, and difficulties in data analysis and service planning. There's a clear need for a way to get quick, actionable feedback on address quality.

The solution: AI-powered address feedback

To address these challenges, we present a solution using Generative AI.

This system is designed to provide users in India with valuable feedback on their addresses, helping them understand potential issues and how to format them better for improved geocoding results and overall accuracy.

The core idea is to use Gemini models on Vertex AI to:

  • Analyze and interpret complex, often malformed, Indian addresses.
  • Identify common errors and inconsistencies.
  • Suggest standardized and corrected versions.
  • Provide clear explanations of the changes made.

The system comes in two form factors: - A REST API - A web UI

How customers in India can use this tool

While the primary goal is to provide feedback on address quality, the benefits extend further:

  1. Improved deliverability: For businesses, understanding how to better structure addresses can mean fewer failed delivery attempts, reduced operational costs, and improved customer satisfaction. Individuals can also ensure they receive their packages and mail more reliably by using well-formatted addresses.
  2. Data enhancement: Companies can use the insights gained from this tool (or integrate the underlying API) to guide the cleaning and standardization of their existing customer address databases, leading to better analytics and targeted services.
  3. Visual verification: The dual-pin map display is particularly valuable. Users can visually confirm if the original and refined addresses point to the same or different locations. This helps identify if the "cleaned" version accurately reflects the intended location or if the original input was too ambiguous or erroneous for correct geocoding.

By understanding the specific challenges in their addresses through the feedback provided, users can take corrective action, update their records, and communicate their locations more effectively.

What is this application?

This web application serves as an interface to the AI-powered address feedback system. It is designed to help users and businesses validate, understand, and improve physical addresses, with a particular focus on the nuances of Indian addresses. The application offers a user-friendly interface where users can:

  • Input an address: Either by typing it directly or by pasting multiple addresses for bulk processing.
  • Receive a cleaned address: The application processes the input and provides a standardized, corrected version based on the AI model's understanding.
  • Understand changes: It highlights the specific modifications made to the original address, offering transparency into the feedback process.
  • Visualize differences: Both the original and the cleaned addresses are pinned on an interactive map, allowing users to visually compare their locations and identify potential discrepancies at a glance.
  • Get detailed components: The geocoded (cleaned) address is broken down into its constituent parts (like street number, locality, city, postal code), providing a structured view.

This app is particularly useful for quickly assessing address quality, understanding potential issues, and seeing how addresses might be better structured for systems that rely on standardized formats.

Backend Architecture: Powered by Gemini and Vertex AI

The intelligence behind this application's ability to understand and refine addresses stems from Google Cloud's advanced AI technologies:

  • **Core address processing: ** The fundamental task of parsing, understanding, correcting, and standardizing address strings is handled by Google's Gemini 2.5 Flash model. When an address is submitted :
    • The frontend application sends the input address to a backend service.
    • This backend service leverages the Gemini API. The Gemini 2.5 Flash model is instructed with a detailed prompt to verify accurate and standardized processing. The core instructions given to the model are as follows:
You are an address cleaning expert. Your task is to take malformed addresses
and output cleaned and standardized versions. All addresses will be from India.

BEGIN:
Follow these instructions:
Remove any mention of "House Number," "H.No," "Door Number," "D.No,"
"Building No", "Flat No." etc. along with the number it's associated with
Remove any "C/O," "S/O," etc.
DO NOT REMOVE any name of building
It should also remove any name of person or actual house numbers etc which
appear after the texts mentioned in the previous point
Ensure there are no duplicate mentions of town names, state names, etc.
If no valid zip code is available, add an error in the Errors field:
"No valid zip code found. Please verify."
Remove mention of any Floors in the address
If there are any mention of "Near or landmark" put that in a new field called
"address_descriptors"
Expand any rd, ln, st and similar other abbreviations to road, lane, street etc.
END:

BEGIN: Structuring the output
Output the cleaned address in a single line.
Output address should put State, Country, Zip code at the end in that order.
If any critical component of the address is missing, mention that in errors section.
**Critically important:** Provide a detailed description of every change made
to the address in the "changes_made" field. Do not omit this field.
IF a House number or unit number was removed add that in a separate field
called "subpremise_details".
Output the errors in the field called "errors". If no errors, provide an empty
array.
Output all responses in JSON format.
END:

This structured prompting guides Gemini 2.5 Flash to:

  • Dissect complex and often unstructured address inputs.
  • Identify and extract key address components (e.g., house /flat number, building name, street, locality, sub-locality, city, state, PIN code).
  • Correct common spelling mistakes and variations.
  • Re-order components into a more standardized format suitable for India.
  • Infer or flag missing critical information where possible.

  • Generate a list of "changes made" and any errors, providing transparency. The model's ability to follow these detailed instructions while handling diverse linguistic patterns and contextual information is key to its effectiveness with varied address formats.

  • Serving and Scalability (Cloud Run on Vertex AI/Google Cloud): The backend service that orchestrates the calls to the Gemini API and returns the results to the frontend is built as a serverless containerized application.

This serverless architecture demonstrates a way to deploy such a service. As a demo application, its primary objective is to allow customers to quickly get some feedback on address quality.

How to use the application

The application is available at India address feedback app.

To use it:

  1. Input your address: Type or paste your Indian address into the input field.
  2. Process the address: Click the "Clean Address" button.
  3. Review the results: The application will display:
    • The cleaned address.
    • A map showing both the original and cleaned locations.
    • A breakdown of the address components.
    • A list of changes made by the AI.
    • Any errors detected.

Direct API call example (for developers)

For developers or systems looking to integrate the address processing functionality directly, the backend service can be called programmatically. Here's an example using cURL:

curl -X POST \
  https://gemini-address-cleaner-480439120941.us-central1.run.app/clean_address \
  -H "Content-Type: application/json" \
  -d '{
    "input_address": "S/O Laum Mirzapur Mirzapur Muzaffarpur Bihar India Mirzapur purani Darbhanga road SELAMBA BIHAR 843103"
  }'

This command sends a POST request with the address string in a JSON payload and will return a JSON response containing the processed address and other relevant information, similar to what's displayed in the application.
This application aims to simplify the complexity of addresses, offering a valuable tool for enhancing accuracy and efficiency, especially in diverse and dynamic environments like India.