Custom RAG Chatbot for Real Estate Financial Analysis

For a Client Building Generative AI tools for Real Estate
‍

The Challenge: Accelerating Investor Due Diligence with RAG

Offering memorandums (OMs) are lengthy financial documents about real-estate properties. Investors usually go through these documents before deciding whether to buy a property or not. The goal of this project was to build a chatbot that can instantly answer complex questions from OMs, making it dramatically easier for investors to find relevant, critical information for due diligence.

Our Solution: Custom RAG Architecture (GPT, LangChain & Pinecne)

We built a chatbot using GPT 3.5 and LangChains. This is a breakdown of our RAG (Retrieval-Augmented Generation) approach:

1. Page-Layout Understanding & Text Extraction

The first step is to extract all text from the OM (which is a PDF document). Since an OM can contain normal text, images, figures, and tables, we need to handle all of these differently. We use an AI-based page-layout understanding module to identify tables, images, paragraphs, etc., separately and then extract text from them, preserving context.

2. Vectorization & Storage (Pinecone)

Once the text is extracted, it is broken down into semantic chunks and stored into Pinecone vector databases.

3. User Query Retrieval (LangChain)

When a user enters a query (e.g., 'what is the projected rent of this property'), we use LangChains to identify the most relevant chunks in the vector database that may contain the answer.

4. Context Augmentation (GPT)

This highly relevant 'context' is then fed to GPT, which uses the source data to answer the user's query with grounding.

5. Hallucination Mitigation

Our prompts for GPT are specifically designed to restrict the model's answers only to the provided context, ensuring the chatbot does not hallucinate and gives accurate and reliable answers.

Results: Reduced Time-to-Decision & Reliable Data Retrieval

Our AI model achieved:

Reduced investor decision-making time by immediately extracting and summarizing relevant financial information.
Highly accurate and reliable answers, avoiding misinformation or incorrect data common with generic LLMs.
Enhanced user experience by allowing them to converse in multiple languages, including through audio messages.

RAG chatbot interface showing summarized real estate financial metrics from an OM PDF. — Overview of the RAG dashboard, summarizing key financial data extracted from the complex Offering Memorandum.

Structured property data output, demonstrating accurate RAG retrieval for due diligence analysis. — Detailed data pane showcasing supplementary property metrics extracted by the AI for investor due diligence.

Custom RAG Chatbot for Real Estate Financial Analysis

See how we built a custom RAG chatbot to analyze Offering Memorandums. Reduce investor research time and eliminate LLM hallucinations.

AI Predictive Modeling to Reduce Meal Plan Rejection & Churn

Stop customer churn. Our predictive AI model identifies meal plans likely to be rejected, reducing rejection rates and protecting brand reputation.

AI Predictive Anomaly Detection for Cold Storage

See how our custom time-series AI model eliminates false alarms, predicts temperature anomalies, and prevents inventory spoilage in cold storage.

Custom RAG Chatbot for Real Estate Financial Analysis

See how we built a custom RAG chatbot to analyze Offering Memorandums. Reduce investor research time and eliminate LLM hallucinations.

Custom Deep Learning for Unstructured Email Data Extraction

See how AxcelerateAI used custom Deep Learning to automate private aircraft charter requests, achieving 98% accuracy & saving dozens of hours weekly.

Custom Computer Vision for Soccer Analytics & Player Tracking

See how our custom Computer Vision models provide real-time soccer analytics, including player tracking, re-ID, and event detection.