Chunking Strategy Advisor

Chunking Strategy Advisor

Purpose

Recommend document chunking strategy: chunk size, overlap, boundaries, and metadata preservation for optimal retrieval quality.

Input: Document types, sample content (optional), query patterns, RAG architecture Output: Chunking specification with parameters, boundary rules, and eval recommendations Examples: See [examples.md](examples.md) for worked input/output.

---

Workflow

Step 1: Analyze document characteristics

| Doc type | Structure | Avg length | Update frequency | |----------|-----------|------------|------------------|

Types: prose, technical docs, code, tables, FAQs, legal, chat logs.

Step 2: Select chunking method

| Method | Best for | |--------|----------| | Fixed-size | Uniform prose | | Semantic | Variable structure, paragraphs | | Recursive | Hierarchical docs (headers) | | Document-specific | Code (functions), tables (rows) | | Agentic | Complex mixed content |

Step 3: Define parameters

| Parameter | Value | Rationale | |-----------|-------|-----------| | Chunk size (tokens) | | | | Overlap (tokens) | | | | Min chunk size | | | | Max chunk size | | |

Step 4: Define boundary rules

• Respect section headers / code blocks / table rows
• Do not split mid-sentence
• Preserve metadata: source, page, section title

Step 5: Recommend eval approach

• Sample 20 chunks for manual review
• Retrieval recall@k on golden queries
• A/B chunk sizes if uncertain

Step 6: Validate

Run Validation checklist.

---

Decision Rules

| Condition | Action | |-----------|--------| | Mixed document types | Use per-type chunking strategies | | Code documents | Chunk by function/class, not token count | | Tables | Chunk by row groups or whole table with summary | | Very short docs (< chunk size) | Keep whole document as one chunk | | No sample content | Use conservative defaults; mark for eval tuning |

---

Validation

• [ ] Document types analyzed
• [ ] Chunking method selected per type with rationale
• [ ] Size and overlap parameters specified
• [ ] Boundary rules documented
• [ ] Metadata preservation defined
• [ ] Eval approach recommended
• [ ] Parameters tied to embedding model context if known

---

Anti-patterns

• One size fits all — same chunk size for code and prose.
• Zero overlap — losing context at boundaries.
• Mid-sentence splits — hurting semantic coherence.
• No metadata — chunks without source attribution.
• Tiny chunks — fragments without context.

---

Best Practices

• Start 256–512 tokens for prose; tune with eval.
• 10–20% overlap for continuity.
• Use header-aware splitting for documentation.
• Store parent document ID and section path in metadata.
• Re-chunk when switching embedding models.

---

Output Structure

# Chunking Strategy: [Corpus Name]

## Document Analysis
| Type | Count | Method |
|------|-------|--------|

## Parameters
| Parameter | Value | Rationale |
|-----------|-------|-----------|

## Boundary Rules
1. [Rule]

## Metadata Schema
| Field | Source |
|-------|--------|

## Eval Plan
[Approach]

## Tuning Notes
[When to revisit]

---

Next Skills

| Outcome | Recommended Skill | |---------|-------------------| | Select embeddings | rag/embedding-strategy-advisor | | Optimize retrieval | rag/retriever-optimizer | | RAG architecture | rag/rag-architecture-designer | | Hybrid search | rag/hybrid-search-advisor |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: patonkikh
• Source: patonkikh/APES
• License: MIT

Install and usage instructions live in the source repository linked above.