Pageindex RAG

Enforcement scripts

This skill includes programmatic enforcement of the catalog-first retrieval invariant:

• scripts/retrieve.py — CSV-based semantic retrieval with authority ranking
• scripts/verify.py — PageIndex MCP tool call plan generator
• scripts/pipeline.py — End-to-end: semantic query → catalog → rerank → verification plan

See scripts/README.md for full usage.

PageIndex Semantic Retrieval Best Practices

Use this skill to answer corpus-grounded questions over the PageIndex library with high recall and high precision.

It is a retrieval-and-verification skill, not a raw indexing skill.

This skill is designed to be standalone. It ships with a bundled catalog snapshot under:

• data/current/

So it can be used remotely in Claude.ai without depending on another local skill installation.

If the user needs a fresher live crawl than the bundled snapshot, then refresh separately with:

• pageindex-reference-index

What "best practice" means here

Use a hybrid workflow:

1. catalog-first retrieval for breadth
2. authority-aware reranking for trustworthiness
3. work / section / topic clustering for semantic coverage
4. live PageIndex page extraction for exact verification
5. citation-rich synthesis in the final answer

Do not jump straight to get_page_content across random PDFs. Do not rely on filenames alone when the catalog already has work_title, topic_category, tags, and relations.

Primary retrieval workflow

1. Check whether the catalog is current enough

Use the bundled snapshot in:

• data/current/

Inspect at least:

• data/current/pageindex_summary.json
• data/current/pageindex_diff.md
• data/current/pageindex_enrichment.md
• data/current/pageindex_work_index.csv
• data/current/pageindex_topic_index.csv
• data/current/pageindex_verification_index.csv

If the user's request depends on the newest live library state and the bundle is stale, note that limitation explicitly and, if available, refresh separately with pageindex-reference-index.

2. Generate candidates from the catalog

Prefer hybrid candidate generation:

• topic
• topic_category
• specific_topic
• work_title
• work_section_title
• source_family
• document_type
• authority
• exam_domain
• hierarchical tags

Remote-first use:

• data/current/pageindex_work_index.csv
• data/current/pageindex_topic_index.csv
• data/current/pageindex_verification_index.csv
• data/current/pageindex_catalog.csv

If you have an execution environment, you may also use bundled query helpers or SQL mirrors, but the skill should remain effective even without them.

3. Rerank by authority and coverage

Default ranking order:

1. official syllabi / examiner reports / guidelines
2. textbook
3. journal
4. reference
5. notes / community-style material

Then balance for breadth:

• prefer multiple source_family values over many hits from one work
• prefer section diversity when synthesizing a topic
• include both full-book anchors and chapter-level hits when useful

4. Verify with live PageIndex only after narrowing the set

Once you have the best candidate documents:

• use get_document to confirm readiness
• for docs >20 pages, use get_document_structure first
• then use get_page_content for tight ranges only

Always preserve exact document / page provenance in your notes.

5. Answer with explicit grounding

Your answer should distinguish:

• metadata-level retrieval
• live page-verified retrieval

When verified, cite:

• document name
• folder path
• page range
• why that source was chosen

Retrieval patterns

Pattern A — "Which texts cover X?"

1. search topic, topic_category, specific_topic, and FTS text
2. cluster by source_family
3. surface 5–10 best works
4. optionally verify the top 2–3

Pattern B — "Compare what different sources say about X"

1. retrieve from at least 3 distinct source_family values
2. prioritize textbooks + official sources
3. verify representative passages live
4. synthesize agreements, disagreements, and level of specificity

Pattern C — "Find the best source to answer X"

1. use verification_index
2. filter by likely topic_category
3. rank by authority, work_title, work_section_title
4. verify the best 1–3 sources live

Pattern D — "Build a reading stack for X"

1. official syllabus / examiner reports first
2. core textbook chapters next
3. supplementary texts after
4. BJA / notes only as adjuncts

Important heuristics

• Prefer work_title and work_section_title over bare filenames.
• Prefer topic_category for recall and specific_topic for precision.
• Use material_kind to separate full books from chapters, front matter, back matter, and problem sets.
• Use source_family to avoid overcounting multiple chapter hits from the same work.
• Use exam_domain to keep ANZCA / CICM filtering intentional.
• Treat pageindex_discrepancies.md as a warning surface, not a footnote.

Rich semantic retrieval checklist

Before finalizing an answer, ask yourself:

• Did I retrieve from the catalog first?
• Did I include the highest-authority sources available?
• Did I avoid collapsing everything into one work?
• Did I verify exact pages when making strong claims?
• Did I cite folder path + page range?
• Did I mention uncertainty if the result is metadata-only?

Recommended references

Read as needed:

• references/retrieval-playbook.md
• references/query-recipes.md
• references/citation-template.md
• references/remote-usage.md
• data/current/README.md

Output contract

For retrieval tasks, usually return:

1. best candidate works
2. best candidate documents
3. verified page-backed evidence where available
4. a brief synthesis
5. suggested next retrieval moves if confidence is partial