Chunking Decision Record: ${CORPUS_NAME}¶

Project: ${PROJECT_NAME}
Document Corpus: ${CORPUS_NAME}
Author: ${AUTHOR_NAME}
RAG Lead / Reviewer: ${REVIEWER_NAME}
Date: YYYY-MM-DD
Status: Proposed | Evaluated | Accepted | Superseded
Related Playbooks: PB-RAG-01-01, PB-RAG-01-02, PB-RAG-01-03

1. Context and Problem Statement¶

Define the corpus characteristics, query patterns, and accuracy goals.

Corpus description: ${CORPUS_SUMMARY} (e.g. ${DOCUMENT_COUNT} engineering manuals, ${FILE_FORMAT} format, average ${AVG_PAGES} pages per document).
Primary failure mode to address: ${PRIMARY_FAILURE_MODE} (e.g. In current fixed-size chunking, ${FAILURE_RATE_PCT}% of retrieval queries return chunks that cut off multi-step procedures or break markdown tables).
Embedding model: ${EMBEDDING_MODEL} (e.g. text-embedding-3-large, default 3072 dimensions, maximum context 8191 tokens; specify ${EMBEDDING_DIMENSIONS}).
Target retrieval latency budget: ${LATENCY_TARGET_P95}.

2. Document Archetype Inventory¶

Classify the corpus into archetypes before benchmarking:

Archetype	Document types / extensions	Document count	Key structural traits
`structured_manual`	PDF, DOCX (SOPs, guides)	${MANUAL_DOC_COUNT}	Hierarchical headings (H1–H4), ordered lists, callout boxes
`tables_and_forms`	PDF (financial reports, specs)	${TABLE_DOC_COUNT}	Multi-column HTML tables, key-value parameter sheets
`long_prose`	PDF, TXT (policy narratives, whitepapers)	${PROSE_DOC_COUNT}	Unstructured prose paragraphs without standard headings
`short_records`	JSON, CSV (support FAQs, product cards)	${RECORD_DOC_COUNT}	Independent records under 1,000 characters each

3. Candidate Profiles Evaluated¶

Record the candidate chunking configurations benchmarked in the sandbox:

version: 1
embedding_model: ${EMBEDDING_MODEL}
candidates:
  candidate_a_fixed_512:
    strategy: recursive
    chunk_size: 2000
    chunk_overlap: 500
    header_levels: 0
    prefix_heading: false
  candidate_b_markdown_headers:
    strategy: markdown_headers
    chunk_size: 2000
    chunk_overlap: 500
    header_levels: 3
    prefix_heading: true
  candidate_c_table_aware:
    strategy: markdown_headers
    keep_tables_whole: true
    chunk_size: 4000
    chunk_overlap: 0
    header_levels: 3
    prefix_heading: true

4. Benchmark Results on Golden Query Set¶

Run evaluation across a representative sample of ${SAMPLE_DOC_COUNT} documents and ${GOLDEN_QUERY_COUNT} labelled queries.

Candidate Profile	Archetype Tested	Recall@5	Recall@10	nDCG@10	Chunk Cut Rate (%)	Median Tokens	Total Chunks
Candidate A (Fixed 2000/500)	`structured_manual`	${CAND_A_MAN_R5}	${CAND_A_MAN_R10}	${CAND_A_MAN_NDCG}	${CAND_A_MAN_CUT_PCT}%	${CAND_A_MAN_MED_TOK}	${CAND_A_MAN_CHUNKS}
Candidate B (Markdown H1-H3)	`structured_manual`	${CAND_B_MAN_R5}	${CAND_B_MAN_R10}	${CAND_B_MAN_NDCG}	${CAND_B_MAN_CUT_PCT}%	${CAND_B_MAN_MED_TOK}	${CAND_B_MAN_CHUNKS}
Candidate A (Fixed 2000/500)	`tables_and_forms`	${CAND_A_TBL_R5}	${CAND_A_TBL_R10}	${CAND_A_TBL_NDCG}	${CAND_A_TBL_CUT_PCT}%	${CAND_A_TBL_MED_TOK}	${CAND_A_TBL_CHUNKS}
Candidate C (Table-aware)	`tables_and_forms`	${CAND_C_TBL_R5}	${CAND_C_TBL_R10}	${CAND_C_TBL_NDCG}	${CAND_C_TBL_CUT_PCT}%	${CAND_C_TBL_MED_TOK}	${CAND_C_TBL_CHUNKS}

5. Decision and Selected Profiles¶

Record the final chosen profiles and the operational rationale.

5.1 Profile Assignment¶

Archetype	Chosen Profile	Chunk Size / Overlap	Extraction Method
`structured_manual`	`structured_manual`	2,000 chars / 500 overlap	Azure Document Intelligence `prebuilt-layout`
`tables_and_forms`	`tables_and_forms`	4,000 chars / 0 overlap	Azure Document Intelligence `prebuilt-layout` (keep tables whole)
`long_prose`	`long_prose`	2,000 chars / 500 overlap	Azure Document Intelligence `prebuilt-layout`
`short_records`	`short_records`	1,000 chars / 0 overlap	Direct JSON/CSV ingestion (1 chunk per record)

5.2 Rationale¶

Heading preservation: Splitting by Markdown headings (#, ##, ###) and prefixing the heading breadcrumb improved Recall@10 from ${BASELINE_RECALL_10} to ${CHOSEN_RECALL_10} on manuals.
Table integrity: Protecting table blocks prevented financial tables from splitting across rows, reducing retrieval hallucination on numerical queries from ${BASELINE_HALLUCINATION_PCT}% to ${CHOSEN_HALLUCINATION_PCT}%.
Cost and storage: Heading-aware chunking generated ${CHUNK_REDUCTION_PCT}% fewer total chunks than fixed-size chunking with 25% overlap, reducing embedding API cost and vector storage.

6. Implementation and Rollback Plan¶

Ingestion pipeline: Implement profile in profile.yaml using the reference chunker in PB-RAG-01-02.
Deployment & Cutover: Re-indexing a live index requires creating a side-by-side index and repointing the index alias without downtime, as detailed in PB-RAG-01-03.
Rollback: In the event of an unpredicted regression in production, repoint the index alias back to ${PREVIOUS_INDEX} in Azure AI Search within seconds.

7. Sign-off and Approvals¶

Role	Name	Decision	Date
Author / GenAI Engineer	${AUTHOR_NAME}	Approved	YYYY-MM-DD
Reviewer / RAG Lead	${REVIEWER_NAME}	Approved	YYYY-MM-DD
Solution Architect	${ARCHITECT_NAME}	Approved	YYYY-MM-DD