Wiki-RAG Ingestion Plan for ECC/SHCC/EGC Atlas

Status

• Workflow layer: 09_ai_workflows.
• Purpose: convert the source-grounded ECC/SHCC/EGC Atlas into a Wiki-RAG / LLM question-answering corpus.
• Critical rule: every answer must be grounded in a source path, a paper card, a source note, or a claim-evidence matrix row.
• Lab-positioning rule: questions about Professor Bang Yeon Lee's research position must search 06_lab_position/ first.

1. RAG purpose

The Wiki-RAG system should help users answer source-grounded questions about ECC, SHCC, EGC, Green ECC, extreme ductility, self-healing, and Professor Lee lab positioning.

The system is designed for:

1. Concept explanation - Example: "What is tensile strain-hardening?" - Priority sources: concept nodes, teaching explainers, foundational matrix.

2. Evidence-backed literature synthesis - Example: "What evidence supports AAS-ECC as Green ECC?" - Priority sources: claim-evidence matrices, source notes, paper cards.

3. Professor Lee lab positioning - Example: "How does Lee lab contribute to recycled selvage fiber ECC?" - Priority sources: 06_lab_position/, lab publication matrix, lab paper cards.

4. Research proposal / manuscript support - Example: "Give a source-grounded novelty claim for low-fiber EGC." - Priority sources: lab positioning cards, global lineage map, claim-evidence rows.

5. Teaching and student guidance - Example: "Explain fiber bridging to a beginner." - Priority sources: 08_teaching/, concept nodes, representative paper cards.

The RAG system must preserve the Atlas evidence hierarchy:

Victor Li book = primary anchor
  -> foundational papers = supporting layer
    -> extension papers = topic-specific extension layer
      -> Lee lab publications = lab positioning layer
        -> teaching and AI workflows = derived use layer

2. Document-type chunking strategy

2.1 Source notes

Source notes live mainly under:

• 00_sources/foundational_papers/source_notes/
• 00_sources/extreme_ductility_extension/source_notes/
• 00_sources/sustainable_ecc_extension/source_notes/
• 00_sources/by_lee_lab_publications/source_notes/

Chunking strategy:

• Chunk by heading level ##.
• Keep frontmatter with every chunk as metadata.
• Keep title, authors, year, DOI, source collection, atlas layer, related nodes, and verification status attached to each chunk.
• Treat ## Key Results, ## Claim-Evidence Candidates, ## Limitations and Cautions, and ## Relationship to Victor Li 2019 Book as high-priority chunks.
• Do not split a claim-evidence table row across chunks.

Recommended chunk size:

600-1200 tokens per chunk, heading-preserving

RAG role:

source note = paper-level evidence summary and interpretation

2.2 Paper cards

Paper cards live under:

• 03_papers/foundational...
• 03_papers/extreme_ductility_extension/
• 03_papers/sustainable_ecc_extension/
• 03_papers/by_lee_lab_publications/

Chunking strategy:

• Usually one paper card can be one chunk.
• If long, split into: 1. Citation + why it matters 2. Main contribution + evidence summary 3. Claim-evidence rows + cautions

Metadata must include:

paper_id:
title:
year:
doi:
source_collection:
atlas_layer:
linked_atlas_nodes:
verification_status:
source_path:

RAG role:

paper card = quick retrieval unit for a specific paper

2.3 Claim-evidence matrix rows

Matrix files:

• 07_visualization/foundational_papers_claim_evidence_matrix.csv
• 07_visualization/chapter2_chapter4_claim_evidence_matrix.csv
• 07_visualization/extreme_ductility_extension_claim_evidence_matrix.csv
• 07_visualization/sustainable_ecc_extension_claim_evidence_matrix.csv
• 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv

Chunking strategy:

• Treat each row as one atomic retrieval unit.
• Do not merge rows unless creating a synthesis answer after retrieval.
• Preserve row-level fields, especially:
• paper_id
• year
• title
• DOI
• atlas_node / atlas_file
• claim
• evidence summary or excerpt
• page / figure / table
• status / verification_level
• mechanism_category
• material_system
• fiber_type
• binder_type
• test_method
• key_metric
• lab_relevance
• priority_for_synthesis

RAG role:

claim-evidence row = highest-precision factual evidence unit

For numeric claims, retrieval should prefer matrix rows over prose nodes.

2.4 Concept nodes

Concept nodes live under:

• 02_concepts/*.md
• selected experiment nodes under 05_experiments/*.md
• selected durability nodes under 03_durability/*.md

Chunking strategy:

• Split by top-level sections.
• Keep node title and linked nodes in each chunk.
• Prioritize chunks named:
• Core idea
• Source basis
• System role
• Anchor evidence
• Supporting sources
• Cautions

RAG role:

concept node = conceptual synthesis and relation map

Concept nodes are useful for explanations, but factual claims should be backed by source notes or matrix rows.

2.5 Material-system nodes

Material-system nodes live under:

• 04_material_systems/*.md

Chunking strategy:

• Split by strategy family or source section.
• For 04_material_systems/green_ecc.md, preserve strategy family sections:
• high-volume fly ash / SCM
• alkali-activated slag
• geopolymer EGC
• recycled fibers / PET / selvage
• rubberized / lightweight / recycled aggregate
• seawater sea-sand / marine resources
• LC3 / low-carbon binders

RAG role:

material-system node = synthesis by material family

2.6 Lab positioning notes

Lab positioning files live under:

• 06_lab_position/our_lab_position_map.md
• 06_lab_position/alkali_activated_ultra_ductile_position.md
• 06_lab_position/recycled_selvage_fiber_position.md
• 06_lab_position/self_healing_position.md
• 06_lab_position/low_fiber_egc_position.md
• 06_lab_position/materials_ai_wikirag_position.md
• 06_lab_position/by_lee_publication_mapping.csv

Chunking strategy:

• Split each positioning card by its 10 sections.
• Preserve status labels such as published, draft, manuscript, and proposal idea.
• Keep reusable paragraphs as separate chunks because they are useful for manuscript and proposal writing.

RAG role:

lab positioning note = strategic interpretation grounded in Lee lab source corpus

Important retrieval rule:

If the question asks about Professor Lee, Lee lab, our research, 우리 연구, 교수님 논문, positioning, novelty, proposal, or manuscript framing, search `06_lab_position/` first.

2.7 Teaching modules

Teaching files live under:

• 08_teaching/teaching_index.md
• 08_teaching/lecture_modules/*.md
• 08_teaching/student_reading_paths/*.md
• 08_teaching/concept_explainers/*.md

Chunking strategy:

• Keep each concept explainer as one chunk.
• Split lecture modules by their ten-section structure.
• Prioritize teaching chunks only for beginner-friendly explanation requests.

RAG role:

teaching module = simplified explanation layer, not primary evidence

Teaching answers must still cite underlying Atlas nodes or evidence rows when factual claims are made.

3. Metadata schema

Recommended metadata fields for every chunk:

chunk_id: unique stable ID
source_path: relative path from ECC_Research_Atlas root
source_type: source_note | paper_card | claim_evidence_row | concept_node | material_system_node | experiment_node | lab_positioning | teaching_module | workflow
atlas_layer: book_anchor | foundational | extension | lab_positioning | teaching | workflow
source_collection: victor_li_book | foundational_papers | extreme_ductility_extension | sustainable_ecc_extension | by_lee_lab_publications | teaching | ai_workflows
paper_id: optional
paper_title: optional
authors: optional
year: optional
doi: optional
journal: optional
atlas_node: optional
atlas_file: optional
mechanism_category: optional
material_system: optional
fiber_type: optional
binder_type: optional
test_method: optional
key_metric: optional
verification_level: verified_from_pdf | verified_from_full_text | verified_from_source_note | pending_user_pdf | book_anchor_seed | candidate
lab_relevance: direct_lab_publication | high_relevance_to_lab_positioning | background_relevance | general_background
priority_for_synthesis: high | medium | low
status: seed | verified | pending_user_pdf | draft | manuscript | proposal_idea
chunk_heading: heading path within document
text: chunk text

For claim-evidence matrix rows, include row-level identifiers:

matrix_path:
row_number:
claim:
evidence_excerpt_or_summary:
page_or_section:
figure_table_equation:

4. Retrieval priority

4.1 General concept questions

Priority order:

1. Concept explainer in 08_teaching/concept_explainers/
2. Concept node in 02_concepts/
3. Primary book anchor / foundational claim matrix
4. Paper cards and source notes for examples

4.2 Numeric or evidence questions

Priority order:

1. Claim-evidence matrix row
2. Source note ## Key Results or ## Claim-Evidence Candidates
3. Paper card evidence summary
4. Full text only if available and needed

4.3 Material-system questions

Priority order:

1. Relevant 04_material_systems/*.md node
2. Claim-evidence matrix filtered by material_system
3. Paper cards/source notes in topic-specific extension folders
4. Lab positioning if the question asks about Lee lab or our research

4.4 Professor Lee lab positioning questions

Priority order:

1. 06_lab_position/our_lab_position_map.md
2. Relevant positioning card: - 06_lab_position/alkali_activated_ultra_ductile_position.md - 06_lab_position/recycled_selvage_fiber_position.md - 06_lab_position/self_healing_position.md - 06_lab_position/low_fiber_egc_position.md - 06_lab_position/materials_ai_wikirag_position.md
3. 06_lab_position/by_lee_publication_mapping.csv
4. 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv
5. Paper cards in 03_papers/by_lee_lab_publications/
6. Global concept/material nodes for comparison

4.5 Proposal or manuscript writing questions

Priority order:

1. Relevant lab positioning card reusable paragraphs
2. 07_visualization/lab_to_global_lineage_map.md
3. Claim-evidence matrix rows with priority_for_synthesis = high
4. Concept/material-system synthesis nodes
5. Paper cards for citations

5. Citation format

Every answer should cite at least one of the following:

1. A claim-evidence row:

[Evidence: 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv, paper_id=alemu_2023_on_crack_healing_in_fiber_reinforced, claim row on permeability reduction]

2. A source note:

[Source note: 00_sources/by_lee_lab_publications/source_notes/alemu-2023-on-crack-healing-in-fiber-reinforced_source_note.md]

3. A paper card:

[Paper card: 03_papers/by_lee_lab_publications/choi-2022-highly-ductile-behavior-and-sustainability_paper_card.md]

4. A concept or material-system node:

[Atlas node: 04_material_systems/green_ecc.md]

Recommended citation format in final natural-language answers:

Claim sentence. (Source: <source_path>; matrix: <matrix_path>, paper_id=<paper_id>)

For numeric values, always cite the claim-evidence matrix row or source note section that contains the number.

6. Answer grounding rule

The RAG system must obey this rule:

No factual claim without a retrievable source path or claim-evidence row.

Answer construction procedure:

1. Retrieve candidate chunks.
2. Separate conceptual chunks from evidence chunks.
3. Use concept nodes for explanation structure.
4. Use claim-evidence rows or source notes for factual and numeric claims.
5. If evidence is missing, state that the Atlas does not yet contain verified evidence for that claim.
6. Never upgrade pending_user_pdf evidence to PDF-verified status.

For lab positioning:

Start with `06_lab_position`, then support with lab claim-evidence rows and paper cards.

7. Hallucination prevention rule

The system must not:

• Invent page numbers, figure numbers, DOI values, or numerical metrics.
• Treat teaching modules as primary evidence.
• Treat source-note-derived claims as PDF verified when original_pdf_status = pending_user_pdf.
• Infer adviser/student lineage or influence unless the source explicitly states it.
• Overstate Professor Lee lab contribution beyond the source-grounded positioning cards.
• Call a material Green ECC if strain-hardening or crack-width control is not preserved or evaluated.
• Treat visual crack closure as equivalent to chloride resistance or corrosion protection.

When evidence is incomplete, answer with:

The Atlas currently supports X from source-note/full-text evidence, but original PDF verification is pending.

or:

I do not find a verified Atlas evidence row for that specific claim.

8. Update cycle

8.1 When new PDFs are supplied

1. Save PDFs in the matching source folder originals/.
2. Match by pdf_filename, DOI, title, and hash.
3. Update source index with: - original_pdf_status = available - original_pdf_path - hash / alias if needed
4. Re-extract or compare full text if necessary.
5. Revalidate claim-evidence rows.
6. Update verification_level.
7. Regenerate missing-PDF request list.

8.2 When a new extension batch is supplied

1. Create a topic-specific folder under 00_sources/<extension_name>/.
2. Preserve full_text, source_notes, paper_cards, and later originals.
3. Create source index.
4. Generate claim-evidence matrix.
5. Add node Supporting sources.
6. Add graph edges.
7. Mark original PDFs as pending_user_pdf if not supplied.
8. Add the extension to retrieval metadata.

8.3 When lab papers are added

1. Ingest under 00_sources/by_lee_lab_publications/ unless a new lab subcollection is explicitly needed.
2. Update 06_lab_position/by_lee_publication_mapping.csv.
3. Add rows to by_lee_lab_publications_claim_evidence_matrix.csv.
4. Update relevant positioning cards.
5. Update graph_lab_to_global_edges.csv.

8.4 Periodic rebuild

Recommended after each major addition:

• Rebuild graph edge files.
• Re-run link validation.
• Regenerate classification summary.
• Update teaching and query templates only if new material changes teaching structure.

9. Example queries and expected source paths

Query 1

Question: What is ECC?

Expected retrieval:

1. 08_teaching/concept_explainers/what_is_ecc.md
2. index.md
3. 02_concepts/strain_hardening_criteria.md
4. 07_visualization/foundational_papers_claim_evidence_matrix.csv

Expected answer style:

ECC is defined by tensile strain-hardening and distributed microcracking, not by fiber addition alone. (Sources: `08_teaching/concept_explainers/what_is_ecc.md`; `02_concepts/strain_hardening_criteria.md`)

Query 2

Question: Why is crack width important for durability?

Expected retrieval:

1. 08_teaching/concept_explainers/why_crack_width_matters.md
2. 05_experiments/crack_width_distribution.md
3. 04_material_systems/self_healing_ecc.md
4. 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv

Query 3

Question: What makes Green ECC different from green concrete?

Expected retrieval:

1. 04_material_systems/green_ecc.md
2. 08_teaching/concept_explainers/green_ecc_explained.md
3. 07_visualization/green_ecc_strategy_map.md
4. 07_visualization/sustainable_ecc_extension_claim_evidence_matrix.csv

Expected answer rule:

State explicitly that greenness without retained PSH/crack-width control is not Green ECC in the Atlas sense.

Query 4

Question: How does Professor Lee's lab contribute to alkali-activated or geopolymer ECC?

Expected retrieval:

1. 06_lab_position/our_lab_position_map.md
2. 06_lab_position/alkali_activated_ultra_ductile_position.md
3. 06_lab_position/by_lee_publication_mapping.csv
4. 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv
5. 03_papers/by_lee_lab_publications/lee-2012-strain-hardening-fiber-reinforced-alkali-activated_paper_card.md

Expected answer rule:

Start from lab positioning, then cite specific lab papers and claim-evidence rows.

Query 5

Question: What is the evidence that recycled selvage fibers can work in ECC?

Expected retrieval:

1. 06_lab_position/recycled_selvage_fiber_position.md
2. 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv
3. 03_papers/by_lee_lab_publications/choi-2022-highly-ductile-behavior-and-sustainability_paper_card.md
4. 03_papers/by_lee_lab_publications/hwang-2025-from-textile-waste-to-high-performance_paper_card.md

Query 6

Question: How should I frame low-fiber EGC novelty?

Expected retrieval:

1. 06_lab_position/low_fiber_egc_position.md
2. 02_concepts/flaw_design.md
3. 03_papers/wang_li_2004_tailoring_preexisting_flaws.md
4. 07_visualization/by_lee_lab_publications_claim_evidence_matrix.csv

Expected answer rule:

Mark low-fiber EGC / EPS bead work as `manuscript / draft` unless final source files have been ingested.

Query 7

Question: Which sources should be used to explain fiber bridging to students?

Expected retrieval:

1. 08_teaching/concept_explainers/fiber_bridging_simple_explanation.md
2. 08_teaching/lecture_modules/03_fiber_bridging_and_interface.md
3. 02_concepts/fiber_bridging_law.md
4. 05_experiments/single_fiber_pullout.md
5. 07_visualization/foundational_papers_claim_evidence_matrix.csv

Query 8

Question: Can AI recommend an ECC mixture from the Atlas?

Expected retrieval:

1. 06_lab_position/materials_ai_wikirag_position.md
2. 09_ai_workflows/atlas_query_templates.md
3. claim-evidence matrices under 07_visualization/
4. Relevant material-system node, e.g., 04_material_systems/green_ecc.md or 04_material_systems/geopolymer_ecc.md

Expected answer rule:

AI may propose candidate directions only as hypotheses. It must cite retrieved evidence and specify required validation tests.

10. Implementation checklist

Before using the Atlas as a RAG corpus:

• [ ] Convert Markdown chunks with metadata.
• [ ] Convert claim-evidence CSV rows into atomic documents.
• [ ] Keep relative source_path in every chunk.
• [ ] Preserve verification_level and original_pdf_status.
• [ ] Route lab-positioning questions to 06_lab_position first.
• [ ] Use teaching modules only for simplified explanations.
• [ ] Require evidence rows for numeric claims.
• [ ] Return citations in every answer.
• [ ] Maintain a missing-PDF and pending-verification filter.