agent-on-capella

An agentic real-estate search experience on Couchbase Capella that personalizes property discovery with vector search and long‑term memory.

Diagram

• Source (agent-on-capella.drawio)

---

Overview

This project demonstrates how Capella’s unified platform enables fast, intelligent, and personalized online buying experiences by combining:

• Vector similarity search for semantic property matching.
• Long‑term memory of user preferences, saved homes, and scheduled tours.
• Tool routing to the best capability per request (LLM, vector index, geospatial, and web search).

Features

• Personalized discovery using embeddings-powered search.
• Long‑term memory of:
  • Available properties
  • Buyer preferences
  • Saved properties
  • Scheduled tours
• Multi‑tool routing for general Q&A, semantic search, geospatial lookups, and real‑time web trends.

Architecture

Single-agent pattern that routes user requests to the optimal tool.

Routes

1. Direct to LLM — us.meta.llama4-maverick-17b-instruct-v1:0 for general questions (e.g., mortgage approval).
2. CB Vector Search Tool — cosine similarity on text embeddings to find relevant properties.
3. Gmaps API / Geospatial Search — uses property latitude/longitude to find nearby schools and restaurants.
4. Realtime Web Search (Tavily API) — current market trends and average prices.

Models

• Text Embeddings: Titan Text Embeddings V2 (1024 dimensions). Similarity: Cosine.
• LLM: us.meta.llama4-maverick-17b-instruct-v1:0.

Vector Index

• Cosine similarity
• 1024 dimensions

Data Model (Couchbase Capella)

• Bucket: properties

  • Scope: 2025-listings
  • Collection: united-states

• Bucket: profiles

  • Scope: buyers
  • Collection: 2025

• Bucket: profiles

  • Scope: tours
  • Collection: 2025

Reference

• Write-up: Search your dream property — building an agentic app on Capella (LinkedIn)