HOME/TECHNIQUE/Retrieval & Grounding/Embedding engineering

TECHNIQUE

Embedding engineering

Retrieval & Grounding

4APPLICATIONS
7OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 7 OPERATORS

Across the pool, embedding engineering is used less as standalone search and more as a vector/similarity stage inside hybrid retrieval, ranking, filtering, and serving pipelines.

Observed Practices

Encode domain objects into embeddings or dense vectors for retrieval, matching, Q&A, recommendations, or clustering: work artifacts, jobs, files, images, structured entities, and journey keywords.

6 of 7 operators
AtlassianCanvaDropboxGrabLinkedInPinterest

Back embeddings with vector databases, vector indexes, FAISS, large-scale indexing, or KV-backed serving paths for fast retrieval and downstream ranking.

4 of 7 operators
AtlassianCanvaGrabLinkedIn

Do not leave embedding or semantic-similarity output as the final decision: combine it with filters, rankers, LLM reranking/answering, classifiers, clustering, diversifiers, or human review.

7 of 7 operators
AtlassianCanvaDropboxGrabLinkedInPinterestUber

Engineer the input context before embedding or retrieval: convert files to raw text and chunks, extract keywords with metadata, or classify intent and fetch external user/profile data instead of embedding only the raw query.

3 of 7 operators
DropboxLinkedInPinterest

Select, evaluate, or adapt embedding models for domain quality, latency, and cost rather than treating the embedding model as fixed.

4 of 7 operators
AtlassianCanvaLinkedInPinterest

Add production mechanics around embeddings where freshness, reuse, or serving latency matters: real-time vector database updates, cached plugin states, nearline embedding publication to key-value stores, and daily incremental inference.

4 of 7 operators
CanvaDropboxLinkedInPinterest

Where Operators Converge

Every observed operator uses embedding or semantic-similarity work as one component in a larger pipeline, with downstream retrieval, ranking, filtering, LLM, UI, serving, or review stages visible in the teardown.

Where Operators Diverge

What gets embedded or compared differs by product domain and granularity.

APPROACH 01

Text/work/search artifacts, jobs, files, security context, or structured entity descriptions are embedded for semantic retrieval, Q&A, or recommendations.

AtlassianDropboxGrabLinkedIn

APPROACH 02

Images are represented as high-dimensional image embeddings for reverse image search.

Canva

APPROACH 03

Keywords extracted from user activity sources are embedded and hierarchically clustered into journey candidates.

Pinterest

APPROACH 04

Generated review comments are compared with a semantic similarity filter to merge overlapping suggestions.

Uber

The post-embedding relevance layer is not standardized.

APPROACH 01

Hybrid retrieval combines traditional fields or strict filters with neural semantic signals, then ranking layers refine results.

AtlassianLinkedIn

APPROACH 02

Vector similarity produces a shortlist and an LLM reranks or answers from the retrieved chunks.

DropboxGrab

APPROACH 03

Embedding search suggestions are shown to designers and can proceed through human review before republishing.

Canva

APPROACH 04

Keyword embedding clusters become journey candidates, then ranking and diversification determine what is surfaced.

Pinterest

APPROACH 05

Semantic similarity is part of post-processing for generated code-review comments, alongside quality scoring and category suppression.

Uber

Model strategy ranges from pretrained/off-the-shelf embedding models to domain adaptation and fine-tuning.

APPROACH 01

Use or select pretrained/off-the-shelf embedding models for the task.

AtlassianCanvaGrabPinterest

APPROACH 02

Fine-tune or domain-adapt representation models for work-shaped or proprietary data.

AtlassianLinkedIn

APPROACH 03

Report strong results without customizing the original embedding approach.

Canva

Watch Items

Cost, latency, and compute pressure materially shape embedding architectures and model choices: operators cite high computational costs, quality/latency/cost balance, time-and-cost reasons for vector database choices, cost efficiency, and added LLM-query latency.

Raw embeddings or raw vector similarity can be insufficient for nuanced or domain-specific retrieval; operators add semantic query understanding, hybrid retrieval, or LLM-assisted ranking, and Grab reports dependence on data quality, complexity, use case, and query patterns.

Evaluation remains a practical constraint: operators use manual nearest-neighbor inspection, LLM-based evaluation or labels, and Grab reports experiments limited to small synthetic datasets with limited queries.

For simpler queries, raw similarity search may still be the efficient option; Grab explicitly warns not to assume the LLM-assisted vector approach is always preferable when computational efficiency matters.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
text-embedding-3-largeservicecommodity
bge-m3libraryestablished
Embedding fine-tuning on domain pairspatternemerging
03

Observed in Production

4 APPS