Skill Embeddings וחיפוש וקטורי ל-Claude

מתי להשתמש

"Embeddings", "Vector search", "Semantic search", "Vector DB", "Pinecone", "Similarity search".

הוראות עבודה

1. What are Embeddings

Numeric representations of text (or images, audio). Similar meanings → similar vectors.

Example

• "dog" → [0.12, -0.45, 0.78, ...]
• "puppy" → [0.14, -0.43, 0.81, ...] (similar to "dog")
• "car" → [0.92, 0.03, -0.31, ...] (different)

2. Embedding Models 2026

3. Choosing Model

Use OpenAI text-embedding-3-small if

• English-primary.
• Cost-sensitive.
• General use cases.

Use Cohere multilingual if

• Hebrew / multilingual content.
• Need consistent quality across languages.

Use Voyage if

• Specific domain (legal, medical, code).

Use open source if

• Privacy critical (self-host).
• High volume + cost matters.

4. Dimensions Trade-off

• More dimensions = better quality, slower search, more storage.
• 1536 dim = sweet spot for most.
• 3072 dim = top quality, more cost.
• 256-512 dim = fast, cheap, less quality.

5. Vector Databases

6. Pricing — Pinecone Example

• Starter: $0 (limited).
• Standard: $70/m (10M vectors).
• Enterprise: Custom.

7. Similarity Metrics

Cosine Similarity (most common)

• 1 = identical, 0 = unrelated, -1 = opposite.
• Good for text.

Dot Product

• Faster computation.
• Some embeddings designed for it.

Euclidean Distance

• Less common for text.
• Good for numerical data.

8. Use Cases Beyond RAG

Semantic Search

• "Find similar products" — not just keyword match.

Recommendation Systems

• "Customers who liked X also liked Y".

Classification

• Find nearest example in labeled dataset.

Clustering

• Group similar items.

Deduplication

• Find near-duplicate content.

Anomaly Detection

• Outlier vectors = anomalies.

9. Sample Code — Semantic Search

from openai import OpenAI
import numpy as np

client = OpenAI()

def embed(text):
    return client.embeddings.create(
        input=text, model="text-embedding-3-small"
    ).data[0].embedding

# Embed corpus once
documents = ["doc1 text", "doc2 text", ...]
doc_embeddings = [embed(d) for d in documents]

# Query
query = "user's question"
query_embedding = embed(query)

# Similarity
similarities = [
    np.dot(query_embedding, d) / (np.linalg.norm(query_embedding) * np.linalg.norm(d))
    for d in doc_embeddings
]

# Top results
top_idx = np.argsort(similarities)[::-1][:5]

10. Hybrid Search (Best Quality)

Combine:
- Semantic (vector similarity)
- Keyword (BM25 / Elasticsearch)
- Weighted scoring

Result: better recall + precision.

Tools

• Weaviate (built-in hybrid).
• Elasticsearch + dense_vector.
• Custom orchestration.

11. Re-embedding Strategy

When to Re-embed

• New embedding model release.
• Major content updates.
• Quality improvements needed.

Don't Re-embed Daily

• Costly.
• Old embeddings still work.
• Schedule monthly/quarterly.

12. Israel Specifics

• Hebrew embeddings: Cohere multilingual best.
• Mixed Hebrew/English: still works, slight quality drop.
• Self-host options for privacy-sensitive Israeli companies.

13. Common Pitfalls

❌ Wrong embedding model for content type. ❌ No reranking — quality plateau. ❌ Embedding too small chunks — loses context. ❌ Mixing embedding models in same DB — incompatible.

14. אסיים בהמלצה.

פרומפט לדוגמה

Embedding model for Hebrew + English mixed corpus?

Pinecone vs pgvector — when each?

Build product recommendation with embeddings.

© 2026 AI Expert Pro | גרסה 1.0.0