Innovative AI Solutions | AI Development, Web & Mobile Apps – Delhi, India
REAL ESTATE · RAG · PROPERTY SEARCH AI

RAG AI That Matches Buyers to Their Perfect Property

We built a conversational property search AI for a real estate portal with 50,000+ listings — buyers describe what they want in natural language and AI instantly matches, compares, and explains relevant properties, increasing enquiry-to-visit conversion by 40%.

40%Better Lead Conversion
3xMore Listings Considered/Visitor
50K+Listings in AI Knowledge Base
28%Longer Session Duration
Build Property Search AI All Case Studies

India's Online Property Search Is Broken for Buyers

India's online property search market is dominated by a handful of major portals — Housing.com, 99acres, MagicBricks, NoBroker — collectively listing over 2 million properties. Despite this enormous inventory, buyer satisfaction with the search experience remains chronically low. Research consistently shows that Indian property buyers visit 6–8 different portals during their search journey, spending an average of 4.5 months from first search to final purchase. The fundamental problem is that all major portals use the same interface paradigm: filter dropdowns for price, location, BHK count, and amenities. This filter-based approach forces buyers to translate their holistic, emotionally complex vision of a home into database query syntax. "I want to be near a good school, commute to Bandra in under 45 minutes, below ₹80 lakhs, somewhere with a view and a good society" cannot be expressed in filter dropdowns — and so buyers spend hours clicking through irrelevant results, frustration builds, and they eventually call a broker who essentially does what AI could do: listen to requirements in natural language and match properties accordingly. Conversion rates for Indian property portals — the percentage of visitors who submit an enquiry — average just 2.1–3.4% industry-wide, an astonishingly low return on the massive SEO and paid traffic investment these portals make. RAG-based conversational search directly solves the discovery problem that drives this low conversion.

🏗️

6–8 Portals Per Buyer Journey

The average Indian property buyer visits 6–8 different portals during their search, spending 4.5 months actively searching. This portal-hopping behaviour is directly caused by search frustration — buyers cannot find what they want on any single portal's filter system and move on. A conversational AI that actually understands buyer requirements creates genuine stickiness, reducing the multi-portal search pattern.

📉

2–3% Industry Conversion Rate

The average visitor-to-enquiry conversion rate across India's major property portals is 2.1–3.4%. Portals invest crores in SEO and Google Ads to drive traffic, then convert fewer than 1 in 30 visitors. The primary cause is search friction — buyers who cannot find relevant listings abandon without enquiring. Even a 40% relative improvement in this conversion rate represents a transformative revenue impact for the portal.

🧠

Brokers Win Because They Listen

Traditional property brokers continue to dominate Indian real estate transactions despite the rise of portals. The reason is simple: brokers conduct a conversation, understand nuanced requirements, and match properties from their mental inventory. RAG-based AI replicates this broker capability at scale — understanding "near a good school, walking distance to the metro, not on a main road, south-facing if possible" and retrieving the most relevant listings from 50,000 options in seconds.

Filter-Based Search Frustrates Buyers

The property portal had invested years in building India's most comprehensive listing database in their regional market — 50,000+ active listings with rich metadata, photographs, builder credentials, and agent contacts. The search infrastructure was technically sophisticated: Elasticsearch-powered filtering, map-based search, saved search alerts, and a recommendation engine based on view history. And yet, conversion was stubbornly low: 28% of visitors who searched submitted an enquiry, and 72% left without any engagement.

User research painted a clear picture of why. Buyers' mental models of their ideal property are holistic and contextual — "I want something near my parents in Andheri East, walking distance from a school for my 7-year-old, but not so close to the main road that it's noisy. My budget is around ₹75 lakhs but I could stretch to ₹85 lakhs for something exceptional. I'd prefer the 4th floor or above for ventilation but not the top floor. Vastu matters to my wife so east or north-facing." This description is perfectly clear to any human who hears it — but it maps to at least 7 separate filter parameters, several of which (proximity to school, relative road noise level, vastu orientation) aren't even available as filters on the portal. Buyers would try to approximate their requirements in the available filters, get irrelevant results, and leave.

Average session depth was 4.2 listings viewed before abandonment — buyers were scrolling a few properties, not finding what they were looking for, and leaving. The support line received 300+ calls per day from buyers who wanted someone to help them find properties matching their requirements — essentially asking staff to do what a conversational AI could handle. Each support call cost the portal an estimated ₹180 in staff time, and the staff who handled these calls had no systematic way to follow up. The portal was paying for human property matching at scale when the entire exercise could be automated and simultaneously made faster, more comprehensive, and available 24/7.

AI That Understands How Real Buyers Think

A RAG-powered property assistant that understands buyer intent, matches listings semantically, and explains why each property fits — like a knowledgeable property advisor who knows every listing in the database.

01

Property Knowledge Base (RAG)

The foundation of the system is a comprehensive vector knowledge base containing all 50,000+ listings, each represented as a rich semantic embedding that captures far more than filter values. Every listing's embedding includes the property description, construction quality indicators from agent notes, neighbourhood lifestyle characteristics, commute time data computed via Google Maps API to 15 major employment hubs, proximity to CBSE and ICSE schools computed from coordinates, builder reputation data from completion history, society amenity profiles, and even listing agent comments about property character ("corner unit with cross ventilation, very bright"). This rich embedding means a query for "near a good CBSE school, not noisy, good ventilation" retrieves listings whose text and structured data actually match those requirements — not just listings that have "school proximity" as a checked amenity. The embedding pipeline uses LlamaIndex with OpenAI text-embedding-3-large, stored in Pinecone for sub-100ms vector retrieval across the full 50K listing corpus.

02

Intent-Based Matching

"Near a good school, 45 minutes from Bandra, below ₹80 lakhs, high floor" — the AI processes this query through a dual pipeline: first, a structured data extraction step that identifies explicit constraints (price ceiling ₹80L, floor preference: high) and implicit requirements (proximity to quality schools, commute time to Bandra under 45 minutes, implied: central or western corridor location); then a semantic vector search across the full listing corpus using these extracted requirements as query terms. The final result set is reranked using a cross-encoder model that evaluates each candidate listing's actual fit against the full query, not just keyword overlap. The top 5 listings are returned with an AI-generated explanation for each: "This 2BHK in Santacruz East is 800 metres from Bombay Scottish School (ICSE), the commute to Bandra Station is 32 minutes by train, it's on the 8th floor, and the asking price is ₹77.5 lakhs." Buyers understand exactly why each result was selected — a level of transparency that filter results never provide.

03

Conversational Refinement

Unlike filter-based search where a refinement requires resetting all filters and re-running the search, the AI maintains the full context of the conversation thread. A buyer who receives initial results and says "Actually I'd prefer Malad West rather than Santacruz, the commute is better for my husband's office in Goregaon" triggers an instant context-update — the AI revises all assumptions based on the new preference without losing any previously established requirements. The conversation can go through multiple rounds of refinement ("can we go down to the 6th floor or below?", "what about 2BHK + study?", "is there anything in a gated community?") with each response building on the complete conversation context. This iterative refinement capability keeps buyers engaged in a productive search experience rather than bouncing between filter resets. Average conversational session depth reached 11.3 exchanges versus 4.2 listing views in the filter-based experience.

04

Enquiry Capture and Nurture

When conversational signals indicate strong buyer interest — typically when a buyer asks specific questions about a particular listing such as availability, possession date, or parking — the AI transitions smoothly to enquiry capture. "You seem quite interested in the Santacruz East apartment — would you like me to connect you with the listing agent and arrange a viewing?" Buyers provide their name and contact details within the chat, and the system simultaneously notifies the listing agent with full conversation transcript and schedules a portal callback. For buyers who show interest but don't immediately enquire, the AI-collected preference data seeds a personalised alert system: email and WhatsApp notifications go out when new listings match their stated requirements, completing the discovery-to-nurture funnel within the portal's own ecosystem rather than losing the buyer to broker channels.

Six-Month Build and Rollout Plan

M1

Listing Database Audit and Schema Standardisation

The first month was devoted to the foundational work that determines RAG quality: auditing all 50,000 listings for data completeness and standardising the listing schema. We discovered that 34% of listings had incomplete or inconsistent location data, 28% lacked agent notes, and only 61% had floor information. We built automated enrichment pipelines: geocoding API calls to standardise coordinates, Google Maps API calls to compute commute times to 15 employment hubs for every listing, and a classification model to infer likely school proximity from coordinates and Google Places data. By month's end, 94% of listings had complete enriched schemas ready for embedding.

M2

LlamaIndex Embedding Pipeline for 50K Listings

Month 2 built and validated the core RAG infrastructure. LlamaIndex document loaders ingested each enriched listing as a structured document with both metadata fields (price, BHK, area, floor) and natural language descriptions synthesised from structured data. OpenAI text-embedding-3-large generated dense vectors for each listing; Pinecone hosted the vector index with metadata filtering capability so price and BHK constraints could be applied as hard filters before semantic ranking — preventing the AI from returning a ₹1.2Cr apartment when the buyer said ₹80 lakhs. The nightly update pipeline was also built in Month 2: new listings added to the portal's PostgreSQL database automatically trigger embedding generation and Pinecone upsert within 1 hour. Sold or withdrawn listings are removed from the index within the same pipeline.

M3

Conversational Search UI (React Chat Widget)

Month 3 built the buyer-facing interface: a floating React chat widget deployable on the portal's existing Next.js frontend with a single script tag. The widget design went through 3 rounds of user testing with property buyers recruited from the portal's existing user base. Key UX decisions: the chat opens with a warm invitation ("Tell me what you're looking for — I'll match it from 50,000+ listings"), not a form. Listing results are displayed as rich cards within the chat — photo, headline specs, and the AI's match explanation — with a "View Full Listing" button and "Enquire Now" CTA. Mobile-first responsive design was critical: 74% of the portal's traffic was mobile.

M4

Commute Time API and School Proximity Integration

Month 4 added the real-time contextual data layers that make AI matching genuinely useful for buyers. Google Maps Distance Matrix API was integrated to compute live commute times from any listing to any buyer-specified destination — buyers can say "I need to reach Whitefield IT Park in under 40 minutes" and the AI retrieves listings where our pre-computed commute data (refreshed weekly) confirms this is achievable. Google Places API integration enabled school proximity queries — the AI knows which listings are within 500m, 1km, or 2km of specific school types (CBSE, ICSE, IB). A proprietary noise level proxy model was also built using road distance data to flag potential noise concerns for listings on major arterial roads.

M5–6

A/B Test, Full Rollout, and Analytics

Month 5 launched a 50/50 A/B test: half of new portal visitors saw the traditional filter-based search; half saw the AI chat widget as the primary search option. After 3 weeks and 45,000 visitor sessions, the AI variant showed 39% enquiry conversion versus 28% for the filter group — a statistically significant 39% relative improvement. Session duration was 28% longer in the AI group, and listings viewed per session tripled. Month 6 completed the full rollout: AI search became the default with traditional filters available as an alternative. A real-time analytics dashboard was built showing AI query trends, most-requested property attributes, and neighbourhood demand patterns — giving the portal's business team unprecedented visibility into buyer intent data they had never previously captured.

Portal Metrics Significantly Improved

🎯

40% Better Conversion

Visitor-to-enquiry rate improved from 28% to 39% — a 39% relative improvement sustained over 6 months post-rollout. Importantly, enquiry quality also improved: the portal's developer clients reported a 28% higher lead quality score (measured by site visit conversion from portal enquiries), indicating that AI is not just generating more leads but more relevant ones. Buyers who articulate their requirements in a conversational search are far better matched to listings they will actually visit and buy.

🏠

3x More Listings Considered

Average listings viewed per session rose from 4.2 to 13.1 — a 3.1x improvement. This matters not just for conversion but for the portal's monetisation model: more listings viewed means more sponsored listings seen, more developer brand exposure, and higher satisfaction scores. Buyers who viewed more listings were also more confident in their eventual enquiry decision, resulting in higher lead quality. The AI introduced buyers to properties in neighbourhoods they hadn't originally considered but which better matched their stated lifestyle requirements — a genuine discovery capability that filter search cannot replicate.

⏱️

28% Longer Sessions

Average session duration grew from 6.8 minutes to 8.7 minutes — meaningful additional time that buyers spend productively engaged with the portal rather than frustrated and bouncing. Long sessions correlate strongly with purchase intent: buyers who spend more than 8 minutes on the portal are 4.2x more likely to submit an enquiry than those who leave in under 3 minutes. The conversational format keeps buyers in an active, exploratory mode rather than the passive scan-and-abandon pattern typical of filter-based browsing. Support line call volume dropped 34% as AI chat handled the discovery assistance previously requiring human staff.

📊

Buyer Intent Data Captured

Every AI conversation generates rich, structured buyer intent data that the portal had never previously captured at scale. AI conversation analytics reveal which neighbourhoods buyers are considering but not finding inventory for (supply gap signal), which lifestyle requirements most frequently go unmet by available listings (product gap signal), at what price points buyers are genuinely active versus aspirationally searching, and which search terms most frequently lead to enquiries. This intelligence now directly informs the portal's inventory acquisition strategy, developer partnership discussions, and editorial content calendar — turning the AI from a conversion tool into a market intelligence platform.

Revenue Impact of Better Conversion

The portal's revenue model is based on paid leads delivered to developers and listing agents — a per-enquiry monetisation structure. With 10,000 monthly unique visitors and a conversion improvement from 28% to 39%, the incremental enquiry volume directly translates to incremental revenue. Additionally, the improvement in enquiry quality (measured by developer lead quality scores) enabled a 12% pricing increase for premium lead products.

MetricBefore AIAfter AIIncremental Value
Monthly unique visitors10,000
Visitor-to-enquiry conversion28%39%+1,100 enquiries/month
Average revenue per enquiry₹2,400₹2,688 (+12% quality uplift)+₹288/enquiry
Enquiry revenue (existing)₹67.2L/year₹112.9L/year+₹45.7L/year
Quality premium revenue uplift+₹50.3L/year
Total incremental annual revenue₹96L/year
System build + Year 1 operating cost₹22 lakhs
Return on Investment4.4x ROI — Year 1

Technologies Used

PythonLangChainLlamaIndexPineconeGPT-4oFastAPIPostgreSQLGoogle Maps APIReact (chat widget)Redis

About This Project

How does the AI stay current as listings are added and sold? +
An automated pipeline runs on a continuous event-driven basis rather than just nightly. When a new listing is added to the portal's PostgreSQL database, a database trigger fires an event to our processing queue. Within 45 minutes, the new listing has been enriched (commute times computed, school proximity calculated, natural language description synthesised), embedded via the OpenAI API, and upserted into the Pinecone index. Sold or withdrawn listings are immediately flagged with a "not available" status in the vector metadata — the AI will never retrieve an unavailable listing as a live result. Price changes and listing updates are also propagated within the same pipeline. For premium developer clients who want their new project launches to appear in AI search within minutes of going live, we offer a priority processing queue with an SLA of under 15 minutes from listing creation to searchability.
Can the AI compare two specific listings side-by-side? +
Yes — direct listing comparison is one of the most-used features in the conversational interface. Buyers can ask "compare the Santacruz East apartment and the Andheri West property you showed me earlier" and the AI generates a structured comparison covering price per square foot (more meaningful than total price for unit size comparison), floor and vastu orientation, commute time to the buyer's specified workplace, proximity to stated school preference, society amenities comparison, possession timeline, builder track record, and the AI's reasoned view on which property better matches the buyer's specific stated requirements. This comparison capability reduces "analysis paralysis" — a common problem where buyers have shortlisted 2–3 properties but cannot decide which to visit first. Buyers who use the comparison feature have a 2.3x higher enquiry conversion rate than those who don't.
How does it handle buyer queries about price negotiation or market trends? +
Market trend queries are handled from a dedicated knowledge base partition containing the portal's published market reports, transaction data aggregates, and neighbourhood price trend analyses updated monthly. For queries like "are prices in Kandivali West going up?" or "is this asking price fair?" the AI draws from actual portal transaction data: "Properties in Kandivali West have averaged ₹12,800 per square foot in transactions recorded on this portal over the last 6 months, with a 4.2% year-on-year appreciation. The listing you're asking about is priced at ₹13,200 per square foot." This factual, data-grounded approach provides genuine value to buyers without overstepping into advice territory. For specific negotiation strategy, the AI recommends speaking with a portal property advisor who can provide personalised guidance — this boundary is carefully maintained.
How does the system handle highly specific queries that only one or two listings match? +
Narrow queries are actually where RAG AI performs best versus filter-based search. When a buyer asks for "a duplex penthouse in Powai with a lake view, below ₹2 crores, available immediately," the AI retrieves from the full 50,000-listing corpus any properties that match all these criteria — even if only 1 or 2 exist. Filter-based search would require the buyer to know the exact combination of filters to apply for "duplex + penthouse + lake view" — which might involve 4–5 different filter interactions. For queries where no listings match all criteria, the AI responds honestly: "I couldn't find an exact match for all your requirements. Here are the closest alternatives — and here's what you'd need to compromise on for each." This honest shortfall response is far more useful than an empty filter result page, and often opens a productive conversation about which requirements are flexible.
Can developers pay to have their listings prioritised in AI search results? +
This was a critical business model question the portal team raised early. Our recommendation — which the portal adopted — is a "sponsored listings" model within AI results: paid developer listings can appear in the top 2 positions, clearly labelled as "Sponsored," provided they meet a minimum relevance threshold for the buyer's query. An irrelevant sponsored listing is never surfaced, protecting the buyer experience and the portal's reputation. Organic results below the sponsored positions are ranked purely on relevance. This model mirrors established patterns from search advertising and has been well-received by both buyers (who understand the convention) and developers (who pay for guaranteed visibility when genuinely relevant). Sponsored listing CPCs in the AI channel have commanded a 35% premium over traditional filter listing placements due to the higher buyer engagement quality.
What data does the AI collect from conversations, and how is buyer privacy protected? +
Buyer conversations are stored with explicit consent disclosed at the start of each chat session. Conversation data is used for: generating personalised listing alerts, improving the AI's matching quality through feedback loops, and aggregated (anonymised) market trend analysis. Individual buyer conversations are never shared with listing agents or developers without the buyer's consent — agents receive only the specific information a buyer has consented to share as part of an enquiry. All data is stored in encrypted PostgreSQL databases within India (AWS Mumbai region) in compliance with IT Act 2000 and applicable data protection standards. Buyers can request deletion of their conversation history at any time. This privacy-first design is both ethically correct and commercially important — buyers will only share genuine requirements with a conversational AI they trust.

Services Used in This Project

Help Property Buyers Find Their Perfect Home with AI

Convert more visitors into enquiries with AI that understands what buyers really want. Get a free demo on your property listings.

Get Free Property AI Demo