Embeddings Builder

Overview

Builds production-grade text embedding pipelines: model selection (OpenAI, Cohere, Voyage, BGE, Snowflake, etc.), batch embedding, normalization, vector database upsert/query patterns, metadata filtering, approximate nearest neighbor tradeoffs, dimensionality reduction for visualization, and complete semantic search examples.

When to Use This Skill

• Adding semantic search to a product (docs, products, messages, code).
• Building recommendation or duplicate detection systems.
• Clustering or topic modeling text data.
• The user says "semantic search", "find similar", "embeddings", or "vector search".

Prerequisites

• Text data to embed.
• Embedding model API key or local model (sentence-transformers, etc.).
• Vector database (Chroma, Qdrant, Pinecone, Weaviate, pgvector, Milvus).

Steps

1. Model selection:

   • Quality vs cost vs latency vs max tokens vs language support.
   • Hosted (OpenAI text-embedding-3-large, Voyage, Cohere) vs self-hosted (BGE, Snowflake Arctic, UAE-Large).
   • Benchmark on your domain if possible (MTEB leaderboard is a starting point).

2. Batch embedding:

   • Always batch (100-500 texts per call depending on model).
   • Handle rate limits and retries.
   • Normalize embeddings to unit length if the DB or similarity metric requires it (most cosine-based do).

3. Vector DB setup:

   • Create collection/index with dimension matching the model.
   • Choose distance metric (cosine, dot product, Euclidean).
   • Design metadata schema (filterable fields: date, source, category, user_id, etc.).

4. Upsert pattern:

   • Generate stable IDs (hash of content + metadata or use your own IDs).
   • Upsert in batches with metadata.
   • Handle updates/deletes (delete old vector, insert new).

5. Query / search:

   • Embed the query with the same model.
   • Top-k + metadata filters.
   • Hybrid search when supported (vector + keyword).
   • Rerank top results with a cross-encoder for better precision.

6. Dimensionality reduction & visualization (optional):

   • UMAP or t-SNE on a sample for exploration.
   • Plotly or matplotlib for 2D/3D scatter with labels.

7. Output:

   • Embedding function (with batching + retry).
   • Vector DB client code for upsert and query (examples for top 2-3 DBs).
   • Semantic search demo script.
   • Normalization and ID strategy notes.
   • Cost/latency estimates for your data volume.

Examples

Complete code for:

• Embedding a corpus of documents with Voyage-2 or OpenAI text-embedding-3-small.
• Upserting to Qdrant and pgvector with rich metadata.
• Performing filtered semantic search.
• A small UMAP visualization of the embedding space. All runnable and with comments.

Edge Cases & Error Handling

• Very long documents: Chunk first (see RAG skill), embed chunks, then aggregate or search at chunk level.
• Multilingual: Choose a multilingual model or separate indexes per language.
• Embedding drift: Re-embed on model upgrades; keep old embeddings for comparison or re-index.

Verification

1. Embed a small set — vectors have the expected dimension and are normalized (if required).
2. Semantic search for a query returns intuitively similar items (manual inspection).
3. Metadata filters work (e.g., only results from 2024).
4. Update a document — old vector is replaced, new search reflects the change.
5. Visualization (if done) shows reasonable clusters.
6. Success: Semantic search returns relevant results with good precision, and the pipeline scales to the target corpus size.

References

• MTEB Leaderboard
• Voyage AI
• Qdrant
• pgvector
• Sentence Transformers
• UMAP