copernicus-rag-core

Full processed core of a 3-tier RAG over all four Copernicus stores — CMEMS (Marine) · CDS (Climate) · ADS (Atmosphere) · EWDS (Early Warning). Private working dataset: originals as markdown, chunks, ready 768-d embeddings, prebuilt Qdrant indexes, all linkage sidecars, and the complete pipeline scripts to rebuild everything from scratch. No PDFs, no images.

Served by the copernicus-rag MCP server (12 tools). Validated 2026-07-23: 50/50 test queries green across all tiers.

Tiers

tier	collection	points	content
L1 discover	copernicus_docs	1,418	dataset cards, all 4 stores
L2 analyze	marine_docs	29,249	CMEMS PUM/QUID/SQO (807 docs / 306 products)
L2 analyze	cds_docs	23,341	CDS/ADS/EWDS PUG/ATBD (766 docs / 165 datasets)
L2 analyze	eqc_qa	1,274	C3S EQC quality reports (74)
L3 method	publications	430,066	12,411 parsed papers, dataset-linked

Layout

originals_md/   cmems/ (813 md) · cds_ads_ewds/ (772) · eqc_reports/ (74)
                notebooks/ (194) · publications_md.tar.gz (11,209 md, MinerU-VLM
                parsed — CC-licensed/PD papers only; 1,202 unlicensed-bronze
                originals removed 2026-07-23, their chunks/embeddings/index
                points remain untouched)
chunks/         per-collection chunks.jsonl + papers.jsonl
embeddings/     *.jsonl.gz — gemini-embedding-2-preview, 768d, L2-norm (ready to load)
indexes/        4 prebuilt embedded-Qdrant dirs (tar.gz) — hybrid dense+BM25,
                publications payloads RELINKED (untar & point the server at them)
metadata/       catalog.json · unified_metadata.json (1,436) · notebooks sidecar ·
                links_by_dataset.json (234 datasets / 8,917 papers / 31,190 links) ·
                publications registry (1,199 DOI) · flagships.json
scripts/        FULL pipeline, per component (see below)
REBUILD.md      full from-scratch rebuild / open-LLM swap guide

Quickstart — plug the RAG database into Qdrant

Two ways, depending on where you want Qdrant to run.

A) Prebuilt embedded indexes (fastest — no re-compute, no server)

The indexes/*.tar.gz are ready-to-serve embedded-Qdrant storage dirs (this is exactly what the MCP server uses). Each tarball unpacks to a qdrant_db/:

pip install "qdrant-client==1.18.0"
hf download dmpantiu/copernicus-rag-core --repo-type dataset \
   --include "indexes/*" --local-dir .
for n in marine_and_cards cds_docs eqc_qa publications; do
  mkdir -p rag/$n && tar xzf indexes/qdrant_$n.tar.gz -C rag/$n
done

dir (after untar)	collections inside	points
rag/marine_and_cards/qdrant_db	marine_docs + copernicus_docs	29,249 + 1,418
rag/cds_docs/qdrant_db	cds_docs	23,341
rag/eqc_qa/qdrant_db	eqc_qa	1,274
rag/publications/qdrant_db	publications	430,066

from qdrant_client import QdrantClient
c = QdrantClient(path="rag/publications/qdrant_db")   # embedded/local mode
print(c.get_collections())                             # -> publications
print(c.count("publications"))                         # -> 430066

Notes:

• Vectors are named: dense (768-d, cosine, gemini-embedding-2-preview, L2-normalized) + sparse (BM25, IDF modifier) → hybrid dense+sparse queries work out of the box. BM25 queries need no embedding model at all (fastembed Qdrant/bm25); dense queries need the same Gemini model (or re-embed — see REBUILD.md).
• Embedded mode holds a single-process lock per dir — one process at a time.
• These dirs are local-mode storage only; you cannot mount them into a Qdrant docker server. For a server, use option B.
• The MCP server (scripts/marine_rag/rag_server.py) expects them at marine_rag/out/qdrant_db, deep_docs/qdrant_db, eqc_qa/qdrant_db, pubs_rag/qdrant_db relative to the repo root.

B) Full Qdrant server (docker / cloud) — see server/GUIDE.md

The server/ folder is a complete, tested deployment kit: docker-compose.yml (Qdrant v1.18) + load_all.py, which downloads the prebuilt indexes and streams all five collections into your server 1:1 — dense + sparse BM25 vectors, relinked payloads, payload indexes; no embedding model or Gemini key needed.

export HF_TOKEN=hf_...
hf download dmpantiu/copernicus-rag-core --repo-type dataset \
   --include "server/*" --local-dir . && cd server
pip install -r requirements.txt
docker compose up -d
python load_all.py --url http://localhost:6333

Full walkthrough (verification, hybrid/filtered query examples, cloud clusters, ops & troubleshooting): server/GUIDE.md.

Note: embeddings/*.embedded.jsonl.gz remain the right starting point when you want to re-embed with a different model (see REBUILD.md); for a faithful copy of the validated database, server/load_all.py is the path — the raw embedding files predate the publication↔dataset relink, the indexes carry it.

Rebuild scripts (scripts/)

Everything needed to regenerate this dataset from the originals — or re-embed with a different model:

marine_rag/       CMEMS: clean_md → chunk_docs → batch_orchestrator (embed) →
                  load_qdrant · cards: build_cds_cards → embed_cds_batch →
                  load_copernicus_docs · rag_server.py (the MCP server itself)
deep_docs/        CDS/ADS/EWDS: fetch_parse → chunk_docs → embed_load
eqc_qa/           EQC reports: fetch → parse → chunk → embed → load ·
                  extract_code + merge_notebooks (notebook sidecar)
pubs_rag/         L3: build_corpus_copernicus → chunk_pubs → embed_orchestrator →
                  load_pubs_qdrant → relink_full → build_links_sidecar
meta_harvest/     01–08: upstream metadata harvest, all 4 stores → unified_metadata
publications/     DOI registry + Crossref/OpenAlex/Unpaywall OA-PDF downloaders
notebook_harvest/ CMEMS gallery / INSTAC notebook parsers
run_test_queries.py   the 50-query validation suite (50/50 pass)
build_bundle.py · build_rag_tree.sh   consolidation helpers

Order for a cold rebuild: REBUILD.md step-by-step; or skip embedding entirely — indexes/*.tar.gz are ready to serve as-is.

Linkage (baked into indexes + sidecars)

• paper↔dataset: 289 via registry (EQC refs) + ~10.4k via flagship-citation map (29 flagship DOIs → verified dataset ids); Qdrant payload linked_products[], flagship_labels[], link_via[], orphan; serve-time sidecar metadata/links_by_dataset.json (built by scripts/pubs_rag/build_links_sidecar.py)
• notebooks attach to dataset cards via sidecar (matched_dataset_id == product_id)
• cards carry n_linked_publications, has_eqc_docs

Notes

• Embeddings: gemini-embedding-2-preview, RETRIEVAL_DOCUMENT, 768 dim, L2-normalized. Query side works with the same model or any 768-d swap after re-embed (see REBUILD.md).
• Licenses (per-paper audit 2026-07-23, OpenAlex×Unpaywall×publisher whitelist, see metadata/publication_licenses.json): of 12,411 papers — 9,813 CC-BY/SA/PD · 1,396 CC-NC/ND · 1,202 no-license/bronze. Full-text originals of the no-license group are NOT stored here (removed; chunks and vectors remain). Copernicus service documents © respective Copernicus services (free use); harvested notebooks retain upstream licenses (incl. some unlicensed training repos). Keep this repo private — NC/ND full texts and unlicensed notebooks are for internal use.