References

Curated references and upstream resources used by this project, grouped by topic. Brief notes explain how each item is used in practice.

Data sources & taxonomies

• Kaggle — https://www.kaggle.com · API/CLI: https://github.com/Kaggle/kaggle-api
  Used in Bronze (cleantech-fetch) to mirror upstream datasets into immutable snapshots.
• OpenAlex Topics — https://docs.openalex.org/api-entities/topics
  Transparent hierarchy (domain → field → subfield → topic) for alignment, coverage, and explainability; OpenAlex Works may be used only for augment‑on‑demand.
• CPC (Cooperative Patent Classification) — EPO overview: https://www.epo.org/en/searching-for-patents/helpful-resources/first-time-here/classification/cpc
  Patent technology vocabulary; kept as per‑chunk metadata for filtering/faceting (e.g., Y02E).
• WIPO Patent Landscape Guidelines — https://doi.org/10.34667/tind.28858
  Anchors for patent‑analytics practice and reporting.

Pipeline & CLI

• Typer — https://typer.tiangolo.com · CLI for ctclean, ctsubsample, ctchunk, ctenrichement, etc.
• Pandas — https://pandas.pydata.org · Tabular processing across Bronze/Silver/Gold.
• Rich — https://rich.readthedocs.io/ · Progress and logging (optional).

Storage & serialization

• Apache Parquet (PyArrow) — https://arrow.apache.org/docs/python/
  Default columnar format; ctclean.io.safe_write handles mixed dtypes and falls back to CSV.gz if needed.
• fastparquet — https://fastparquet.readthedocs.io/ · Alternate engine.

Retrieval, ranking & RAG

• BM25 (Robertson & Zaragoza) — overview: https://www.nowpublishers.com/article/Details/INR-019
  Sparse lexical baseline used by python -m bm25.
• Sentence‑BERT (bi‑encoder) — https://arxiv.org/abs/1908.10084 · docs: https://www.sbert.net/
  Dense retrieval and semantic chunking.
• Reciprocal Rank Fusion (RRF) — https://doi.org/10.1145/1571941.1572114
  Robust fusion of BM25 and dense ranks in the Hybrid adapter.
• Cross‑encoder re‑ranking — https://www.sbert.net/examples/applications/cross-encoder/README.html
  Improves early precision before composition.
• Maximal Marginal Relevance (MMR) — https://dl.acm.org/doi/10.1145/290941.291025
  Reduces redundancy and improves topical breadth.
• RAG (survey) — https://arxiv.org/abs/2409.14924 · Primer on RAG methods and evaluation.

Graph & communities

• Neo4j Cypher — LOAD CSV — https://neo4j.com/docs/cypher-manual/current/clauses/load-csv/
  Used to ingest GraphRAG CSV exports.
• GraphRAG (patterns & repo) — https://github.com/microsoft/graphrag
  Query‑focused summarization and graph‑assisted retrieval ideas for exploration.
• Leiden communities — https://www.nature.com/articles/s41598-019-41695-z
  Community detection used for cluster labels & diversity analysis.

Vector & lexical stores

• ChromaDB — https://docs.trychroma.com · Dense vector store for python -m dense.
• rank‑bm25 (PyPI) — https://pypi.org/project/rank-bm25/ · Sparse retriever used by python -m bm25.
• NLTK — https://www.nltk.org/ · Stopwords/tokenization utilities (optional).

Evaluation

• pytrec_eval — https://github.com/cvangysel/pytrec_eval · TREC metrics for the benchmark runner.
• Introduction to IR — https://nlp.stanford.edu/IR-book/ · Textbook reference for nDCG, MAP, etc.

LLM providers (optional)

• Google Gemini — https://ai.google.dev/ · Embeddings/summaries (GEMINI_KEY).
• OpenAI — https://platform.openai.com/docs · Optional summaries/rerankers (OPENAI_KEY).

Documentation stack

• MkDocs — https://www.mkdocs.org · Material for MkDocs — https://squidfunk.github.io/mkdocs-material/
• mkdocstrings (Python) — https://mkdocstrings.github.io/python/ · mkdocs‑jupyter — https://github.com/danielfrg/mkdocs-jupyter
• mkdocs-video — https://github.com/soulless-viewer/mkdocs-video

Study & internal docs

• GRAINS preliminary study (Gerber, 2025) — internal PDF in this repo.
  Source for acceptance budgets (Recall/nDCG/Precision/Coverage, explainability ≥ 2 citations, p95 latency envelopes), single‑adapter serving path, and evaluation protocol used across these docs.