# Building Biz Ask Anything: From Prototype to Product

[Yelp](https://yomu.fyi/company/yelp) · Maria Christoforaki, Group Tech Lead; Shree Shalini Pusapati, Software Engineer · Mar 27, 2026

**Type:** Problem & solution

## Summary

Yelp expanded Yelp Assistant to business pages by developing a question-answering architecture that extracts concise, evidence-backed answers from reviews, photos, and structured facts. The production system replaces static snapshots with near-real-time streaming pipelines for reviews and structured properties alongside weekly batch ingestion for menus and website data. For each incoming query, the system evaluates intent, retrieves relevant context through keyword and embedding searches, and constructs targeted prompts with dynamically selected few-shot examples. Synthesized answers are augmented with supporting visuals and citations before being streamed token-by-token to users. Generating suggested questions directly from specific business content rather than generic categories improved user engagement by roughly fifty percent and reduced unanswerable queries by twenty-six percent.

## Context

Users visiting Yelp business pages often need quick, direct answers to specific questions without sifting through large volumes of reviews, photos, menus, and structured business details. The original 2024 Yelp Assistant only matched users with service professionals and lacked the ability to answer business-level questions.

## Approach / What changed

Yelp evolved a prototype into Biz Ask Anything, an LLM pipeline backed by near-real-time streaming indices, an Entity-Attribute-Value Cassandra fact store, and a sub-100ms content-fetching API. Incoming queries undergo safety checks, intent analysis, retrieval via keyword and photo embedding search, prompt composition with semantically matched few-shot examples, and token streaming.

## Takeaways

- Yelp adopted an Entity-Attribute-Value Cassandra schema for structured business facts, enabling flexible attribute additions without schema migrations since the LLM consumes unstructured strings.
- Data ingestion streams reviews and business facts from source-of-truth databases within 10 minutes, while website, menu, and community content are updated via weekly batch jobs.
- Generating suggested questions from business-specific reviews and summaries rather than broad categories resulted in a ~50% engagement lift and a ~26% drop in unanswerable queries.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [Data Pipelines](https://yomu.fyi/topic/data-pipelines), [LLMs](https://yomu.fyi/topic/llm), [Search](https://yomu.fyi/topic/search), [Streaming](https://yomu.fyi/topic/streaming)

[Read original post](https://engineeringblog.yelp.com/2026/03/building-baa-from-prototype-to-product.html)