OLX Magic turns product search into a conversational experience that can refine requests, compare listings, search from images, and handle broad shopping goals.
2
The agent uses a model router, search tools, guard rails, streaming responses, and hybrid retrieval across OLX's catalog.
3
OLX found that users often prefer familiar search controls, while agentic systems still create problems with latency, cost, scalability, and monetization.
Summary
Nishikant Dhanuka and Beatriz Ferreira describe OLX Magic, a shopping assistant for OLX's secondhand marketplace. It accepts natural-language requests, images, and other forms of input, then uses agentic flows to search the catalog, ask clarifying questions, compare items, and suggest related products. The speakers show two product versions. The first centers on chat, while the second combines a familiar search bar with agent features such as Magic helper, instant results, smart groups, contextual tags, and natural-language refinement. Under the hood, the system uses commercial and fine-tuned models, a model router, web, text, and visual search tools, and hybrid keyword-semantic retrieval. OLX's user research led the team to reduce reliance on chat and onboarding cards. The speakers are direct about the remaining issues: agent responses take seconds instead of hundreds of milliseconds, multiple model calls raise costs, tools must scale with marketplace traffic, and the team still needs a workable monetization model.
OLX Magic lets shoppers refine a request over several turns
OLX Magic is designed for shopping requests that change as people see products. A shopper might begin by asking for size 41 shoes, then switch to size 42 or adjust the budget after seeing options. The assistant supports that refinement through conversation and can suggest follow-up questions. It can also compare selected products in a side-by-side table and add a written recommendation. Nishikant Dhanuka presents this as a way to help shoppers make decisions without forcing them to express the full request in one search.
The assistant handles broad requests as well as specific product searches
Magic can receive a narrow request, such as a particular Dyson model, or a broad goal such as finding things for a camping trip. For a broad request, it asks clarifying questions and then returns groups of products, such as tents and other camping gear. Beatriz Ferreira gives another example, finding a birthday gift for someone's mother. The agent creates a plan, asks questions until it has enough information, and then makes parallel calls to text and visual search so it can present several relevant options.
User research pushed the product toward familiar search controls
The first version relied heavily on a conversational interface, but interviews showed that chat was not suitable for every search. Someone searching for a yellow jersey may want immediate results rather than follow-up questions. The team also found that few customers used onboarding cards, even though those cards were familiar to ChatGPT and Gemini users. In the second version, OLX moved toward a powerful search bar with familiar controls. It added a Magic helper, instant results while slower responses load, sorting, smart groups, and visible buttons such as Smart Compare.
Natural-language filters can expose information missing from structured fields
The second version lets a shopper type refinements such as a location and a maximum price, then presents them like ordinary filters that can be removed. It can also search for attributes that are not defined as catalog filters. A shopper might request a smartwatch with a heart-rate monitor or ECG function. Because OLX is a secondhand marketplace, sellers do not always fill in every field, so the relevant detail may only appear in a listing description or image. Magic tries to use that unstructured information while keeping the interaction familiar.
The agent connects several models and search tools behind one request
Beatriz Ferreira describes an architecture in which the agent interprets the request, creates a plan, calls tools, and returns the result to the user. A model router connects both commercial models and internally fine-tuned models. Available tools include web search, text search, visual search, and URL parsing. Text and visual search include retrieval and ranking flows. The system also has language detection and guard rails. Responses are streamed back because the agent may need to pass through several calls before it can answer.
Hybrid search combines exact terms with semantic meaning
The speakers describe search quality as central to a shopping assistant. OLX Magic combines keyword search with semantic search. Keyword matching can miss relevant products when the listing does not contain the exact query terms. Semantic search can find products with similar meaning, while hybrid search combines both signals. The example uses a query for an aquamarine blouse. Keyword search returns only a product with matching words, while semantic search can find blouses with similar blue colors. The combined approach is used to ground the assistant's results in the catalog.
Visual search can apply a change to an image-based request
Magic supports visual search from an uploaded product image. A shopper can upload a dress or toy car and ask for similar items. The request can also include a modifier, such as changing a yellow toy car to red or changing another product attribute. The agent keeps the image as the basis for the search and applies the requested modification. Beatriz Ferreira says this can cover attributes such as color, shape, sleeve length, and category-specific properties.
Adoption remains split between natural-language queries and ordinary keywords
OLX is testing Magic by placing an entry point inside the existing search experience. The user's search context is passed into Magic, which helps avoid a cold start and gives the assistant an initial prompt. Nishikant Dhanuka says some users begin writing more context in natural language, while many continue to use the assistant like a keyword search box. The product therefore has to support both behaviors. The team is changing the interface while users are also learning how to use the new technology.
Agentic search has unresolved latency, cost, scaling, and monetization problems
The speakers report that agent responses can take five to ten seconds, while conventional search is expected to respond in roughly 300 to 500 milliseconds. Agents also make many model calls, which increases cost even as individual LLM calls become cheaper. Every tool used by the agent must scale to marketplace traffic. Guard rails require ongoing testing because users find unexpected ways to use the product, and false positives can block normal actions such as uploading an image without text. OLX is also still working out how to monetize the new experience.
"The long tail is almost like that is going to take up a majority of the work."Demetrios Brinkmann27:41
Who should watch
You are building an AI shopping or discovery experience on top of a large, messy marketplace catalog.
Your team is deciding whether a chat-first interface fits users who already understand conventional search.
You need practical examples of agent architecture, hybrid retrieval, visual search, evaluation, guard rails, and the operational cost of multiple model calls.
