Enhancing E-commerce Search with Vector Embeddings and Generative AI

mercado_libre_/_ · E-commerce · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Marketplace semantic search

Mercado Libre needed search to understand complex natural-language shopping intent rather than relying only on word matching against item titles and attributes.

Constraints:

The platform serves a very large Latin American commerce and payments user base.
Traditional word matching works for simple queries but fails for complex intent.
Search is a core business pathway because many successful purchases start with a query.
Complex queries need to connect buyer intent to seller inventory even when exact words do not match.
The source is a Google Cloud showcase transcript with some promotional framing.

Architecture

Mercado Libre represents each item as a vector embedding and stores those embeddings in Google Cloud Vector Search. At query time, the system embeds the user's natural-language query, asks Vector Search for the closest matching item embeddings, and uses those semantic matches to improve result listings for complex queries that word matching misses.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  user-search-query["User search query"] --> query-embedding-service["Query embedding service"]
  item-embedding-pipeline["Item embedding pipeline"] --> google-vector-search["Google Vector Search"]
  query-embedding-service --> google-vector-search
  google-vector-search --> search-ranking-and-listing["Search ranking and listing"]
  search-ranking-and-listing --> buyer-and-seller-feedback["Buyer and seller feedback"]
  buyer-and-seller-feedback --> item-embedding-pipeline

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`User search query`	Natural-language or long-tail buyer intent entered in Mercado Libre search.	`Query embedding service`
`Query embedding service`	Converts the user's query into a vector representation.	`Google Vector Search`
`Item embedding pipeline`	Generates vector embeddings for item representations.	`Google Vector Search`
`Google Vector Search`	Stores item embeddings and retrieves semantically close matches to query embeddings.	`Search ranking and listing`
`Search ranking and listing`	Builds the result page from semantically matched items and existing ranking signals.	`Buyer and seller feedback`
`Buyer and seller feedback`	Clicks, conversions, and seller-buyer matching signals that indicate search quality.	`Item embedding pipeline`

Data flow

Item content is embedded and inserted into a Google Vector Search database. When a user submits a query, the query is embedded, Vector Search finds nearby item embeddings, and the resulting matches improve the ranked listing for complex shopping intent.

“we generate a vector embedding for represent each item and then insert all these embeddings in a Google's Vector search database”transcript, stated

“when a users makes a query what we do is we compute the embedding for that users's query and ask to Vector search to F find the closed matching items”transcript, stated

Why this architecture

Mercado Libre moved beyond word matching because complex shopping requests can describe intent without sharing words with the right item titles. Embeddings let the search system compare item and query meaning, improving result quality for queries like gift ideas or task-oriented computer searches.

“traditionally our search engine relies heavily on war matching”transcript, stated

“with this approach we can increase the quality of the rle listing especially for these complex queries”transcript, stated

Models used: vector embeddings · Google Cloud Vector Search · generative AI

Scale

Mercado Libre has 144 million unique users whose experiences need contextualization.
“that's 144 million unique people whose experiences need to be contextualized understood served at any given point in time”transcript, stated
Unique active users grew by over 50% year over year.
“the number of unique active uh users on their platform has grown by over 50% year overy year”transcript, stated
Most successful purchases start with a search query.
“most of the successful purchases start with a query made by the user in our search bar”transcript, stated
About half of distinct user queries are complex.
“like half of this distance queries that user made are complex ones”transcript, stated

Alternatives considered (roads not taken)

Word matching against titles and attributes.

Rejected because: It fails for complex natural-language shopping requests that do not share terms with the right products.

“this approach works really well for lot of saries but if the user want for look for a more complex query”transcript, stated

Only exact Samsung-and-TV-style matching.

Rejected because: The system needed to understand intent such as gifts for a soccer fan, not just title word overlap.

“we try to find all the items with the word Samsung and the word TV in the items title or attributes”transcript, stated

Patterns

`rag-retrieval` (retrieval)

Mercado Libre retrieves semantically close item embeddings for user query embeddings.

“ask to Vector search to F find the closed matching items related to that query embedding”transcript, stated

`retrieval-ranking` (retrieval)

Vector matches improve the ranked listing quality for complex shopping queries.

“we can increase the quality of the rle listing especially for these complex queries”transcript, stated

`query-rewriting` (retrieval)

The system translates natural-language intent into an embedding query that can retrieve products without exact word overlap.

“understanding the user intent understanding what they're trying to search for”transcript, stated

Tradeoffs

Search representation

Chose Embeddings for items and queries. over Word matching alone..

Rationale: Embeddings support complex queries where titles and attributes do not share exact words with the user request.

“to solve this problem we start using embeddings for item representation”transcript, stated

Business target

Chose Improve complex-query quality. over Optimizing only simple exact-word searches..

Rationale: Half of distinct queries are complex and improving them can lift overall click and conversion rates.

“if we can improve the quality for these complex queries the the results quality for these complex queries and then we can increase the overall click rate a conversion rate”transcript, stated

Gotchas (what bit them)

Word matching produced irrelevant results for complex intent.

Root cause: Exact title or attribute overlap could not infer that a soccer fan gift should retrieve soccer-related merchandise.

Resolution: Represent items and queries as embeddings and retrieve semantically close items.

“no soccer related items no messy merchandise and this is because traditionally our search engine relies heavily on war matching”transcript, stated

Search quality affects both buyers and sellers.

Root cause: Bad semantic matching prevents buyers from finding items and sellers from connecting to buyer needs.

Resolution: Use vector retrieval to improve matching for complex queries.

“this helps our buyers to find what they need but also this allows our sellers to connect more efficiently with the buyer needs”transcript, stated

Evaluation

The transcript reports business evaluation through click rate, conversion rate, and qualitative examples of complex-query result quality rather than a formal benchmark.

Metrics: Click rate · Conversion rate · Complex-query result quality

“we can increase the overall click rate a conversion rate by a significant amount”transcript, stated

Outcomes

Improved result-listing quality for complex queries.
Connected buyers and sellers more efficiently when exact words did not match.
Positioned search to better handle long-tail gift and intent-based queries.
Aimed to increase click rate and conversion rate.

Mercado Libre cites 144 million unique people, over 50% year-over-year user growth, and roughly half of distinct queries being complex.