NammaHubballi replaced traditional category browsing with an AI query engine powered by Google Gemini and pgvector. Users ask natural language questions, and the system retrieves grounded answers exclusively from a curated, high-quality local business database.
The Context
Local business discovery in mid-sized Indian cities is broken. Google Maps shows chains and hotels. JustDial has outdated listings and aggressive ad injection. The real local economy — the specific tailor who does alterations, the diagnostic lab that opens on Sundays, the chartered accountant who handles GST returns — is invisible online. The gap was a directory that was genuinely useful for local queries, not just comprehensive in listing count.
The Approach
Built the discovery layer as an AI query engine rather than a search bar. Users type natural language questions — 'Where can I get a bike serviced near Vidyanagar?' — and the system returns grounded answers from the local business database using Gemini with pgvector RAG. The business listings themselves are structured with opening hours, service categories, and verified contact information — curated manually for quality, not scraped for quantity.
Grounding AI to prevent hallucinated business info.
Grounding Gemini responses to only the local database — preventing hallucinated business details while still allowing natural language flexibility. The solution was a strict RAG architecture: Gemini only generates from retrieved context, never from training data, with a fallback 'not found in directory' response for out-of-database queries.
4,000+ businesses indexed. The AI query layer handles 60%+ of all discovery sessions — users prefer the natural language interface over the category browse by a 3:2 margin. Architecture replicated for AskBLR in 2025.
Engagement Meta
Technology Context
Quick Navigation
FRONTEND
API LAYER
SERVICES
DATABASE
INFRA
Architectural Decisions
Why this specific stack was chosen over standard defaults for this build.
The AI is only as good as the data underneath it.
A city AI product built on scraped or low-quality data will confidently give wrong answers. Curated data at lower volume outperforms scraped data at high volume every time.
Hallucination prevention is architecture — not prompting.
Prompt instructions to 'only use provided data' are not reliable at scale. A strict RAG architecture where the model cannot generate outside the retrieved context is the only production-safe approach.
Discovery quality beats listing quantity.
A directory with 500 well-curated, accurate listings is more valuable than one with 10,000 outdated or duplicate entries. The quality of each listing determines whether the user comes back — not the count.
The Final Result
4,000+ businesses indexed. The AI query layer handles 60%+ of all discovery sessions — users prefer the natural language interface over the category browse by a 3:2 margin. Architecture replicated for AskBLR in 2025.
The AI is only as good as the data underneath it.
A city AI product built on scraped or low-quality data will confidently give wrong answers. Curated data at lower volume outperforms scraped data at high volume every time.
Hallucination prevention is architecture — not prompting.
Prompt instructions to 'only use provided data' are not reliable at scale. A strict RAG architecture where the model cannot generate outside the retrieved context is the only production-safe approach.
Discovery quality beats listing quantity.
A directory with 500 well-curated, accurate listings is more valuable than one with 10,000 outdated or duplicate entries. The quality of each listing determines whether the user comes back — not the count.