Embedding Strategy Advisor

Purpose

Select and configure embedding models and indexing strategy for a RAG corpus based on domain, language, and quality requirements.

Input: Corpus description, languages, domain (general/technical/legal), budget, latency requirements
Output: Embedding strategy document with model recommendation, dimension trade-offs, and index configuration Examples: See examples.md for worked input/output.

Workflow

Step 1: Profile the corpus

• Languages and multilingual needs
• Domain specificity (general, code, medical, legal)
• Corpus size (chunks count)
• Query-document similarity type (semantic, keyword-hybrid)

Step 2: Evaluate embedding options

Compare API models (OpenAI, Cohere, Voyage) vs open-weight (BGE, E5, GTE).

Step 3: Recommend embedding model

Selection criteria:

• Retrieval quality on domain sample (MRR, recall@k)
• Latency for query embedding
• Cost at projected volume
• Self-hosting feasibility

Step 4: Define index configuration

• Distance metric: cosine, dot product, L2
• Index type: HNSW parameters (ef, M)
• Quantization if corpus > 1M chunks

Step 5: Plan re-embedding strategy

When corpus or model changes:

• Version embeddings
• Batch re-embed pipeline
• Dual-index migration period

Step 6: Validate

Run Validation checklist.

Decision Rules

Validation

• Corpus profile complete
• ≥2 embedding options compared
• Recommendation with quality/cost/latency rationale
• Index configuration specified
• Re-embedding strategy documented
• Dimension and metric choices explained
• Eval method for embedding quality stated

Anti-patterns

• Default model — using text-embedding-3-small without domain eval.
• Dimension mismatch — mixing embeddings from different models in one index.
• No re-embed plan — model lock-in without migration path.
• Ignoring query embedding cost — only indexing cost considered.
• Over-quantization — precision loss without quality check.

Best Practices

• Eval embeddings on domain-specific golden queries.
• Match embedding model to chunking strategy.
• Store model version in index metadata.
• Benchmark query latency at p95, not average.
• Align with chunking-strategy-advisor output.

Output Structure

# Embedding Strategy: [Corpus Name]

## Corpus Profile
| Attribute | Value |
|-----------|-------|

## Options Compared
| Model | Dims | Quality | Cost | Latency |
|-------|------|---------|------|---------|

## Recommendation
**Model:** [name]
**Rationale:** [why]

## Index Config
| Setting | Value |
|---------|-------|

## Re-embedding Plan
[Strategy]

## Eval Results
| Metric | Score |
|--------|-------|

Next Skills