Optimal RAG configuration differs by discipline: STEM subjects perform best with 512-character chunks and temperature 0.3 (faithfulness 0.97), while humanities require 1000-character chunks and temperature 0.1 (context recall 1.00, faithfulness 0.98)¶
Claim · All claims
Evidence · 1 study (2 entries) · designr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? For STEM subjects the optimal chunk size was 512 characters with model temperature 0.3, achieving RAGAS faithfulness of 0.97. → Anna Ostrowska 2026
q2 i? For humanities, larger chunks (1000 chars) and lower temperature (0.1) achieved context recall of 1.00 and average faithfulness of 0.98. → Anna Ostrowska 2026 (2)
Evidence¶
Anna Ostrowska 2026¶
Anna Ostrowska, Michał Kukla, Gabriela Majstrak, Jan Opala, Sebastian Pergała, Jan Skwarek, Anna Wróblewska. (2026). From Surface Learning to Deep Understanding: A Grounded AI Tutoring System for Moodle. https://tinyurl.com/mrytc2dr doi:10.24963/ijcai.2026/984
q2 · i? · design · r2
RAGAS (LLM-as-a-Judge) evaluation of the system's retrieval and generation pipelines across disciplines. The technical optimization experiments found the printed STEM configuration, with "achieving a RAGAS faithfulness of 0.97" for technical formulas.
"for STEM subjects: The optimal chunk size was 512 characters, and the best performance was reached with a model temperature of 0.3 were optimal for technical formulas, achieving a RAGAS faithfulness of 0.97"
Anna Ostrowska 2026 (2)¶
Anna Ostrowska, Michał Kukla, Gabriela Majstrak, Jan Opala, Sebastian Pergała, Jan Skwarek, Anna Wróblewska. (2026). From Surface Learning to Deep Understanding: A Grounded AI Tutoring System for Moodle. https://tinyurl.com/mrytc2dr
q2 · i? · design · r2
The same RAGAS evaluation of the retrieval and generation pipelines reported the humanities configuration, "achieving a context recall of 1.00, and an average faithfulness of 0.98" with the larger chunks and lower temperature.
"for humanities: larger chunks (1000 chars) and a lower temperature (0.1) were required to capture interpreted meanings, achieving a context recall of 1.00, and an average faithfulness of 0.98"
Discussion¶
Related Claims¶
- Temperature-construct interactions: constructs with moderate theoretical coherence benefited from higher temperatures, while well-defined constructs required deterministic settings — related
- Gemini Flash 2.0 was selected for production over GPT-4o-mini and Llama 3.1 for its superior faithfulness and speed-to-cost ratio — related
- NCERT-tailored RAG preprocessing and figure-aware indexing improve retrieval precision over a generic-chunking baseline, per the authors' assertion — related
- Differences in AI-supported instruction effects across STEM disciplines are not statistically significant
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
rag-configuration-discipline-dependent-moodle-tutor |
title |
Optimal RAG configuration differs by discipline: STEM subjects perform best with 512-character chunks and temperature 0.3 (faithfulness 0.97), while humanities require 1000-character chunks and temperature 0.1 (context recall 1.00, faithfulness 0.98) |
description |
Optimal RAG configuration differs by discipline: STEM subjects perform best with 512-character chunks and temperature 0.3 (faithfulness 0.97), while humanities require 1000-character chunks and temperature 0.1 (contex... |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
2 — anna-ostrowska-2026, anna-ostrowska-2026-2 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.