Service
Retail and e-commerce AI on Google Cloud
Formare AI builds AI for retailers and e-commerce platforms on Google Cloud: multimodal product search with image, text and voice on Vertex AI Vector Search, re-ranking that uses live stock, price and click signals, and merchandising agents on Google ADK that choose substitutes and react to catalog events. Our reference grocery search build, a hackathon entry, is configured for under 300 ms P99 retrieval and over 95% Recall@20 across about 10,000 products (design targets from Google's benchmarks, not production results). Projects start from €8,000 or at €600–€900 per day, after a free 30-minute discovery call.
Updated
Scope
What we build for retailers
Customers rarely know the name of the product they want. They look for "the pasta with the yellow label" or "something like this photo", and keyword search returns nothing for either. We build the search, the ranking and the merchandising logic around that problem.
- Multimodal product search: a photo, a description or a voice request finds the closest products in the catalog
- Re-ranking that combines semantic relevance with live stock, price, margin tier and click-through rate
- Substitution engines that propose in-stock replacements in the same category and price band
- Merchandising agents on Google ADK that interpret a request or a catalog event and return a structured decision
- Product knowledge graphs built from BigQuery catalog data, with relationships such as SUBSTITUTABLE_FOR, SAME_BRAND and BOUGHT_TOGETHER
- Brand search optimization and a voice shopping layer on the Gemini Live API
Architecture
How an AI product search pipeline fits together
| Stage | Google Cloud service | What happens | Reference figure (target, not measured in production) |
|---|---|---|---|
| Embed | Vertex AI multimodal embeddings | Products with an image get an image-plus-text embedding; the rest fall back to text-only | About 90% of the catalog had images; 10,000 products embedded per hour |
| Retrieve | Vertex AI Vector Search | Approximate nearest-neighbour search returns the 150 closest candidates | Target: under 300 ms P99, over 95% Recall@20 |
| Enrich | Vertex AI Feature Store | Inventory status, price, margin tier and click rate are attached to each candidate | Target: under 2 ms per lookup |
| Re-rank | Vertex AI Ranking API | Candidates are re-scored for relevance to the query | Google-reported 15–25% precision gain over raw vector distance |
| Decide | Google ADK on Cloud Run | A planner agent runs the steps above in parallel and assembles the answer | Built on Gemini 2.5 Flash |
| Speak | Gemini Live API | Real-time voice in and out over a WebSocket, on the same pipeline | Optional layer |
The index setting that matters most is the number of approximate neighbours: 150 is the value Google's benchmarks associate with over 95% recall, while lower values such as 50 are faster but miss relevant products.
Merchandising
Choosing substitutes when a product is out of stock
When stock runs out, a substitution agent takes the vector-search candidates and keeps only those in the same category, in stock and within 15% of the original price. The Ranking API score sorts what is left, with a small bonus for items that are in stock.
The rules are ordinary code, not a prompt, so a merchandiser can read them, test them and change them without retraining anything.
Cost and timeline
How long retail AI takes and what it costs
Most engagements run 2–12 weeks. The main drivers are catalog size, how many products have usable images, and which live signals (stock, price, clicks) are available to feed the ranking.
Pricing is €600–€900 per day, from €8,000 for a fixed-scope project paid in milestones, or from €4,500/mo for a retainer. Rates are in euros and VAT may apply. If you are not sure where to start, the 4-Week AI Strategy & Architecture Audit (from €8,000) delivers a written architecture plan prioritised by ROI. The 30-minute discovery call is free.
Limits
What the figures do and do not tell you
The latency, recall and precision figures above are design targets taken from Google's published benchmarks and documentation, as set out in the linked build articles. They are not production measurements and not client results: the multimodal shopping assistant is a hackathon entry, built for Google's Gemini Live Agent Challenge. Results on your catalog depend on its size, image coverage, data quality and traffic, so we measure on your data before promising a number.
If keyword search with good synonyms and filters already meets your goals, we say so on the discovery call.
Work
Retail builds, and how each is used
Hackathon entry
Multimodal AI Shopping Assistant
Gemini Live Agent Challenge — Hackathon
Grocery product search on a catalogue of about 10,000 products, built for the Gemini Live Agent Challenge. Image + text search on Vertex AI embeddings and Vector Search, in three variants: text, image, and live audio over WebSocket with the Gemini Native Audio API.
- <300ms P99 retrieval (target)
- >95% Recall@20 (target)
- +15–25% precision (Google-reported)
Not production measurements. The latency and recall figures are targets for this index configuration, taken from Google's published Vector Search benchmarks. The precision gain is the figure Google reports for the Ranking API.
How it was built
Reference build
Senior Merchandiser AI Agent
Retail catalog intelligence — multi-agent
Multi-agent merchandiser: Gemini planner, parallel vector search and re-ranking, Feature Store enrichment.
- +15–25% search precision (Google-reported)
- Real-time Feature Store <2ms (target)
Reference figures, not client measurements. The precision gain is the one Google reports for the Ranking API; the Feature Store latency is a design target.
How it was built
Reference build
BigQuery Knowledge Graphs
Graph construction + NL → Cypher agent
Turns flat BigQuery product tables into a Neo4j knowledge graph, queried in plain English via NL→Cypher.
How it was built
FAQ
Questions about Retail and e-commerce AI development
Cost, timelines and fit, answered before the first call.
What is multimodal product search?
Multimodal product search lets customers search with a photo, a description or their voice instead of an exact product name. Products and queries are turned into embeddings, vectors that place similar items close together, and the nearest products are retrieved. Our reference build uses Vertex AI multimodal embeddings, so a photo and a phrase like "the pasta with the yellow label" can find the same product.
How fast and how accurate is AI product search on Google Cloud?
Our grocery search build, a hackathon entry on a catalog of about 10,000 products, is configured for under 300 ms at P99 and over 95% Recall@20 in Vertex AI Vector Search, the figures Google's published benchmarks report for that setup. Google reports a 15–25% precision gain from re-ranking with the Vertex AI Ranking API. These are design targets, not production measurements. Your numbers will depend on catalog size, image coverage and traffic, so we measure on your data before promising any.
Can search use our stock, price and click data?
Yes. Signals such as inventory status, current price, margin tier and click-through rate are read from Vertex AI Feature Store at query time and combined with semantic relevance, so a product that has just gone out of stock drops in the ranking without manual work.
How does the system pick substitutes for out-of-stock products?
A substitution agent takes candidates from vector search and keeps only those in the same category, in stock and within 15% of the original price. The Ranking API score is the main sort key, with a small bonus for in-stock items. The rules are plain code, so a merchandiser can read and change them.
How long does a retail AI project take and what does it cost?
Most projects run 2–12 weeks, depending on catalog size, image coverage and which signals are available. Pricing is €600–€900 per day, from €8,000 for a fixed-scope project paid in milestones, or from €4,500/mo for a retainer. Rates are in euros and VAT may apply. If you are unsure where to start, the 4-Week AI Strategy & Architecture Audit (from €8,000) ends in a written plan. The 30-minute discovery call is free.
Can customers shop by voice?
Yes. The voice variant of our shopping assistant uses the Gemini Live API for real-time speech in and out over a WebSocket, on top of the same search pipeline. Our article on Gemini Live covers the audio handling and the configuration mistakes that stop it working.
Read next