Scope
- Lesson file:
docs/projects/multihop-wikipedia-qa/index.md (id multihop-wikipedia-qa, slug /projects/multihop-wikipedia-qa, _category_.json position 25)
- Example folder
examples/multihop-wikipedia-qa/:
data/articles/*.md — six self-written Wikipedia-style articles crafted so a few questions need facts from two articlesdata/test_questions.json — six bundled test questions, three genuinely multi-hop, each with an expected answermain.py — CLI with --rebuild, --query, --provider, --top-k; single-hop vs multi-hop pipelines printed side by side with evidence chunksnotebook.ipynb — Colab/Kaggle/Binder-ready, corpus embedded, getpass API-key cellpyproject.toml, uv.lock, .env.example (copied from agentic-code-reviewer), .gitignore, .python-version, README.md
- Registration files:
docs/projects/index.mdx, src/data/projects.ts (prepended entry, date 2027-08), src/pages/index.tsx (HomepageProjectCard with <Translate>)
Framing note
Builds on the RAG App project's single-hop pipeline. The pedagogical core is iterative retrieval: round one retrieves, the model judges sufficiency, and if the evidence is incomplete the model writes a follow-up query for a second round — then both paths print the exact evidence chunks used, so a wrong answer comes with an audit trail. Corpus and test questions are crafted so the multi-hop failure is reproducible and visible, and the related-projects section links only to on-main slugs.
Scope
docs/projects/multihop-wikipedia-qa/index.md(idmultihop-wikipedia-qa, slug/projects/multihop-wikipedia-qa,_category_.jsonposition 25)examples/multihop-wikipedia-qa/:data/articles/*.md— six self-written Wikipedia-style articles crafted so a few questions need facts from two articlesdata/test_questions.json— six bundled test questions, three genuinely multi-hop, each with anexpectedanswermain.py— CLI with--rebuild,--query,--provider,--top-k; single-hop vs multi-hop pipelines printed side by side with evidence chunksnotebook.ipynb— Colab/Kaggle/Binder-ready, corpus embedded,getpassAPI-key cellpyproject.toml,uv.lock,.env.example(copied from agentic-code-reviewer),.gitignore,.python-version,README.mddocs/projects/index.mdx,src/data/projects.ts(prepended entry, date2027-08),src/pages/index.tsx(HomepageProjectCard with<Translate>)Framing note
Builds on the RAG App project's single-hop pipeline. The pedagogical core is iterative retrieval: round one retrieves, the model judges sufficiency, and if the evidence is incomplete the model writes a follow-up query for a second round — then both paths print the exact evidence chunks used, so a wrong answer comes with an audit trail. Corpus and test questions are crafted so the multi-hop failure is reproducible and visible, and the related-projects section links only to on-main slugs.