# RAG vs Fine-Tuning Decision Matrix

Canonical: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/
Publisher: Pharos Production
Published: 2026-08-18
Method version: 2026-08-18.1
Language: English

## Direct answer

RAG is a candidate when a workload needs changing external knowledge or inspectable source passages and an authoritative corpus plus retrieval operating path exist. Fine-tuning is a candidate when the primary gap is stable task behavior and representative permitted examples, an evaluation set, training operations, and rollback exist. A hybrid is a candidate when both needs are independently evidenced and both lifecycles can be operated. If a required data asset, evaluation contract, or operating capability is missing, the method returns baseline first, blocked, or insufficient evidence.

## Important boundaries

- RAG can preserve document and chunk provenance but does not guarantee claim-level citation support.
- Fine-tuning changes behavior or task specialization; it is not a dependable live store for changing facts.
- Privacy, cost, and latency depend on the concrete provider, topology, traffic, quality target, and operations.
- No universal dataset-size, document-count, accuracy, latency, cost, or performance threshold is published.
- Every result is a candidate for a bounded experiment, not a production guarantee.
- Research synthesis and implementation were AI-assisted. Independent human technical review is not claimed.

## Public artifacts

- Decision tool and architecture guide: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/
- Methodology and source ledger: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/methodology.html
- Criteria: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/criteria.json
- Rules: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/rules.json
- Claims: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/claims.json
- Sources: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/sources.json
- Scenarios: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/scenarios.json
- Comparison CSV: https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/data/decision-matrix.csv

## Citation guidance

Cite the exact claim or rule ID, method version, checked date, and canonical page. Recheck vendor and provider statements after 2026-08-18. Treat examples as decision patterns rather than measured customer outcomes.