Leading Commercial Real Estate Tech Launches AI Lease Chatbot Powered by Nexla

Background

The customer is a commercial real estate software platform used by portfolio managers to manage leases across multiple properties. Historically, the platform provided deterministic analytics and data extraction capabilities. Lease data lived across two systems: a Snowflake table containing structured lease records, and an Amazon S3 bucket holding the underlying PDF lease documents, which can run 50 or more pages each.

Portfolio managers needed to answer questions about individual leases without manually reading through lengthy documents, and to run cross-portfolio analyses without writing custom queries. The platform identified a retrieval augmented generation (RAG) chatbot as the solution, and selected Nexla as the data infrastructure layer.

The Challenge

Building a production RAG system for commercial real estate required solving three distinct problems:

• Combining structured data from Snowflake with unstructured data from PDF documents into a single, query ready system
• Chunking and embedding PDF content in a way that preserved lease specific context, so semantic search could surface accurate clause-level answers
• Supporting two distinct query modes: single lease lookups tied to a specific document, and portfolio wide aggregations requiring calculations across many leases simultaneously

Direct database or S3 access by the agent was not viable. Any production system needed a governed, consistent data layer that the agent could rely on for both retrieval and calculation.

The Solution: Nexla as the Data Foundation

The platform built its RAG chatbot with Nexla handling all data movement, transformation, and access. The agent never queries Snowflake or S3 directly. All data flows through Nexla datasets, providing a single governed interface for the retrieval layer.

Data Pipeline Architecture

Two pipelines feed the system:

• Snowflake lease data: Structured records flow into Nexla datasets, making tabular lease information available for portfolio-level calculations and filter queries.
• PDF lease documents: An S3-to-chunking pipeline runs inside a Nexla flow. Documents are retrieved from S3, processed through a chunking model that breaks

The chunking pipeline follows the sequence: S3 bucket source, chunking model within the Nexla flow, vector database sink. This keeps the entire document processing workflow within Nexla, with no external orchestration required.

Retrieval and Answer Generation

When a user submits a query, the agent retrieves relevant content from both data sources using vector embedding. Rather than keyword matching, the system uses semantic similarity, meaning varied phrasings of the same question return consistent results. The agent then analyzes the retrieved evidence and returns a response with citations pointing back to the source document or record.

The platform provided an equation book for complex real estate metrics. These formulas are applied at query time for calculations such as base rent per square foot ranges, weighted average lease terms by sector, and square footage at risk in a given expiration window.

Use Cases

Single-Lease Queries (Document View)

When a portfolio manager opens a lease inside the the platform portal, the chatbot appears on the right side of the document view. The manager can ask questions about that specific lease without scrolling through the full document.

Example question types covered in the 163-question test set include:

• Tenant improvement allowance and base rent for a specific tenant at a named property
• Security deposit and prepayment terms
• Signage rights and permitted use clauses
• CAM charges, tax estimates, and HVAC responsibilities
• Late payment interest rates and administration fees
• Landlord demolition and redevelopment rights

Portfolio-Wide Queries (Homepage Interface)

From the the platform portal homepage, users access an Express-style interface for portfolio-level questions. These queries cross multiple leases and require the agent to aggregate, filter, and calculate across the full dataset.

Example question types covered in the 98-question multi-lease test set include:

• Which leases expire in a given calendar year, and what is the combined square footage and rent at risk
• Which tenants have renewal options or contraction options, and when are notice deadlines
• Percentage rent clauses and applicable sales thresholds across the portfolio
• Sector breakdown by square footage and number of leases
• Base rent per square foot range across the portfolio, including highest and lowest paying tenants
• Exclusive use rights and potential conflicts between tenants in the same portfolio

Sample Validated Questions

The customer supplied a structured test set of 261 questions to validate system accuracy before launch. A representative sample is shown below.

Why Nexla

Nexla’s role in the RAG system extends beyond a simple connector. The platform handles:

• Data access governance: The agent queries Nexla datasets rather than raw Snowflake tables or S3 paths, maintaining a consistent and auditable access layer.
• Pipeline orchestration: The S3-to-vector-database chunking workflow runs as a Nexla flow, with no separate orchestration infrastructure.
• Multi-source unification: Structured and unstructured data from two different systems are made available through a single interface the agent can use uniformly.
• Scalability: Nexla’s platform, which processes over 10 trillion records per year and offers 600+ connectors, provides the throughput and connectivity required for a production enterprise deployment.

The chatbot launches to end users through the platform portal, available behind login across both the document view and portfolio homepage interfaces. The system was validated several hundred representative test questions covering single-lease document lookups and complex portfolio-wide calculations before release.

Portfolio managers can now ask natural language questions about any lease or across their full portfolio, receiving answers with citations, without writing queries or reading through multi-page documents manually.