Data Science: CompStak’s Agentic Search

SKIP AHEAD TO

• How Agentic Search Works
• How the Agents Understand CRE Data
• Product Impact

Introduction

As the commercial real estate (CRE) industry increasingly invests in niche sectors, the need to pinpoint exact, relevant information has outgrown the capabilities of traditional label- or keyword-based search. Conventional search systems rely on pre-defined categories and exact text matching, which often fail to keep pace with the fast-changing focus of CRE investments. As a result, they can overlook valuable results when terminology shifts, or surface irrelevant records when the same keywords appear in unrelated contexts.

Agentic semantic search addresses these challenges by leveraging AI to understand the intent and meaning behind a user’s query, rather than just matching words. This enables the system to return results that capture nuanced concepts—such as a tenant’s business context or specific space usage—making lease data more discoverable and actionable.

For instance, determining whether a retail property is “anchored” by a grocer or a luxury tenant can be ambiguous and subjective (e.g., is Uniqlo considered a High-End or Aspirational Retailers ?). By enabling leases to be searched semantically through natural language, users can find the most contextually relevant data first, and then refine results by their specific use case.

How Agentic Search Works

To deliver a truly intelligent search experience, CompStak has implemented an agentic search framework layered on top of a Retrieval-Augmented Generation (RAG) architecture.

In this approach, search is no longer a single static retrieval step. Instead, an AI agent orchestrates the search process actively reasoning through a sequence of retrieval and synthesis steps to ensure the best possible answer. When a user submits a query, the agent:

1. Interprets Query Intent – The system goes beyond literal keyword matching to understand the semantic meaning and business context behind each query, automatically translating industry terminology and concepts into precise data retrieval strategies. Advanced Semantic Understanding: The agent recognizes that “FIRE leases in Midtown Manhattan” requires multiple layers of interpretation:
   • Industry Acronym Recognition: “FIRE” = Financial, Insurance, and Real Estate sector tenants
   • Geographic Mapping: “Midtown Manhattan” = Specific CompStak submarkets (Midtown East, Times Square, Grand Central, Penn Station, etc.)
   • Business Context: Understanding that the user wants leases occupied by companies in these specific industries within those geographic boundaries
2. Iterative Retrieval – The system intelligently combines multiple search approaches to deliver the most relevant results. Rather than relying on a single search method, the agent uses both structured data filtering and semantic understanding to cast a wide net, then automatically ranks results by relevance to ensure the best matches appear first. Smart Search Strategy:
   • Precise Filtering: Applies exact criteria like location, property type, and lease terms to narrow the dataset
   • Contextual Matching: Uses AI to understand the business meaning behind your query, finding leases that match your intent even when the exact words differ
   • Automatic Ranking: Results are intelligently ordered so the most relevant leases appear at the top, saving you time scanning through less relevant matches
   This approach ensures you get comprehensive results that are both accurate and meaningful to your specific search needs.
3. Context Assembly – Selecting and combining the most useful context from multiple sources, ensuring both breadth and depth of coverage.
4. Generative Synthesis – Passing the curated context into a large language model to produce data search results that answer users need.
5. Conversational Refinement – The system maintains context across multiple queries within a search session, enabling users to iteratively refine their search scope without repeating previous criteria. This allows for natural, conversational interactions where users can progressively narrow or expand their search parameters. Example in Practice:
   • Initial Query: “New York City cold storage locations over 10,000 sqft”
   • System Response: [Returns results across all NYC boroughs]
   • Refinement: “I mean Manhattan specifically”
   • System Understanding: Automatically applies the geographic refinement while preserving all other search criteria (cold storage, 10,000+ sqft), delivering Manhattan-specific results without requiring users to restate their full requirements.

How the Agents Understand CRE Data

Understanding User Queries Semantically

CompStak’s agentic search begins by interpreting the true intent behind each user query, going beyond exact keyword matches. This capability is strengthened through agent’s prompt built upon CompStak’s deep domain expertise, built from processing over two million crowdsourced lease comps enriched with diverse commercial real estate ontologies. This extensive dataset allows the system to recognize industry-specific terminology and contextual nuances, ensuring search results align with the meaning, not just the wording, of a query.

Enriching Tenant Data with LLM Embeddings

To further enhance relevance, CompStak’s tenant information is transformed into high-dimensional embeddings using large language models (LLMs). These embeddings capture not only factual attributes, but also the business context, industry positioning, and space usage of each tenant. As a result, searches can connect queries with the most contextually aligned tenants, enabling users to discover insights that traditional metadata or classification systems might overlook.

Product Impact

The agentic search is powering CompStak’s AI Chat feature on our Enterprise platform

The search is also available through API, which can even be sent and processed in Excel with Power Query or any dashboard tool that supports API integration.