RAG vs Fine-Tuning Decision Matrix: Versioned Rules and Scenarios

This software package preserves version 1.0.0 of a deterministic decision aid for planning an evaluation of retrieval-augmented generation (RAG), fine-tuning, a hybrid system or an existing prompting baseline. It also returns blocked and insufficient-evidence states when the recorded conditions do not support an architecture candidate.

Pharos Production, an AI software development company, developed the RAG vs fine-tuning decision matrix to make the assumptions behind that choice inspectable: every decision rule links to named claims and sources.

The archive contains the complete source tree from commit 6ea24ef75c3ad09a3bb76d1c20a173bfab592559, including the static browser application and its tests. The canonical registries define 12 criteria, 11 decision rules, six advisory rules, 15 claims and eight sources. Four illustrative scenario cards describe intended uses. A separate reproduction supplement provides six complete input fixtures derived from the existing tests, expected outputs and a command that checks each result. These examples exercise the rule engine. They contain no model responses or measured model-performance data.

How the decision works

Each criterion takes yes, no or unknown. The engine requires an answer for every registered criterion and rejects invalid values or unregistered criterion identifiers. It evaluates decision rules in ascending priority order and selects the first matching outcome. Advisory rules add follow-up work without changing that outcome.

The six output states are RAG, FINE_TUNING, HYBRID, BASELINE_FIRST, BLOCKED and INSUFFICIENT_EVIDENCE. The decision export includes the winning rule, recorded inputs, triggered advisories, disqualifiers and provenance identifiers. The export records registry versions so a decision can be read against the rules that produced it.

For example, a workload may need changing facts while also having suitable examples for behavior adaptation. If its authoritative corpus is explicitly absent, the highest-priority rule returns BLOCKED. A plausible fine-tuning path does not remove that knowledge dependency. Unresolved core answers and a missing evaluation set have their own rules. Priority order matters when several conditions apply together.

The companion RAG vs fine-tuning architecture guide discusses knowledge freshness, behavior adaptation and hybrid systems. It provides company-authored background; its broader performance and cost statements are not thresholds in this rule engine.

The four scenario cards in the source registry are teaching examples with partial evidence patterns. They must be completed before evaluation. The six executable fixtures in the supplement instead specify every input and cover all six output states.

Reproduce and inspect

Use Node.js 22 or newer. The project has no runtime or development dependencies. From the extracted package root, run node reproduce.mjs to compare the six fixtures with their saved expected results. In source/, run npm test, npm run check:data and npm run check:release to verify the original tests, registry references and public data mirrors. The checks use local files. They do not call a language model. Instructions and per-file SHA-256 checksums accompany the archive.

Scope and reuse

The output identifies an evaluation candidate. It does not establish that an architecture will satisfy a particular workload's quality target, permissions, latency budget or operating requirements. The software assigns no probability or confidence score and defines no universal accuracy, cost or dataset-size threshold. Teams must evaluate passage support, behavior regressions and recovery in their intended deployment. The related RAG development services from Pharos Production describe implementation work spanning retrieval, evaluation, permission boundaries and ongoing monitoring. Those deployment activities require evidence beyond this matrix's architecture recommendation.

The source registries retain their original verification date of August 18, 2026. Package checks on September 14, 2026 concern source integrity and deterministic behavior. They do not refresh every external claim. Research synthesis and implementation were AI-assisted. Independent human technical review is not claimed.

Code and its accompanying software documentation retain the MIT license. Original structured data uses CC BY 4.0. The license manifest identifies the scope of each license. Referenced third-party material retains its own terms. CITATION.cff supplies authorship, version and repository metadata. No DOI is asserted before a record is assigned one.