server/GUIDE.md

# Copernicus RAG — full Qdrant server setup

Stand-alone guide: from zero to a running **Qdrant server** carrying all five
collections of `dmpantiu/copernicus-rag-core`, byte-identical to the validated
indexes (dense 768-d + sparse BM25 vectors, relinked payloads, payload indexes).
No embedding model, no Google/Gemini key, no re-computation involved.

## What you get

| collection | points | content | key filterable payload fields |
|---|---|---|---|
| `copernicus_docs` | 1,418 | L1 dataset cards, all 4 stores | `product_id`, `doc_type`, `store` |
| `marine_docs` | 29,249 | CMEMS PUM/QUID/SQO chunks | `product_id`, `doc_type`, `chunk_type`, `section_path` |
| `cds_docs` | 23,341 | CDS/ADS/EWDS PUG/ATBD chunks | `dataset_ids`, `store`, `doc_type`, `doc_url` |
| `eqc_qa` | 1,274 | C3S EQC quality reports | `dataset_id`, `store`, `doc_type`, `aspect` |
| `publications` | 430,066 | 12,411 parsed papers, dataset-linked | `doi`, `paper_id`, `journal`, `year`, `domains`, `orphan`, `linked_products`, `chunk_type` |

Every collection has **named vectors**: `dense` (768-d, cosine,
`gemini-embedding-2-preview`, L2-normalized) and `sparse` (BM25, IDF modifier)
— so keyword search works with no embedding model at all, and hybrid search
works if you can embed queries (see below).

## 1 · Prerequisites

- Docker + docker compose (any recent version).
- Python 3.10+.
- Access to the private HF dataset (ask the owner to add you as a collaborator
  on `dmpantiu/copernicus-rag-core`) and a HF access token
  (huggingface.co → Settings → Access Tokens → read).
- Disk: ~2.5 GB download + ~6 GB Qdrant storage. RAM: 2–4 GB is plenty
  (payloads are stored on disk).

## 2 · Get this folder & install deps

```bash
export HF_TOKEN=hf_...           # your token
pip install -U huggingface_hub
hf download dmpantiu/copernicus-rag-core --repo-type dataset \
   --include "server/*" --local-dir .
cd server
pip install -r requirements.txt
```

## 3 · Start the Qdrant server

```bash
docker compose up -d
curl http://localhost:6333/readyz     # -> all shards are ready
```

Dashboard: <http://localhost:6333/dashboard>. Data persists in
`./qdrant_storage` across restarts. To protect the server, uncomment
`QDRANT__SERVICE__API_KEY` in `docker-compose.yml` first.

## 4 · Load all five collections

```bash
python load_all.py --url http://localhost:6333
```

What it does: downloads the four `indexes/*.tar.gz` from HF (≈2.5 GB, cached in
the temp dir), untars them, opens each prebuilt index locally and **streams the
points into your server** — vectors, payloads and payload indexes are copied
1:1 from the validated build (50/50 test queries green).

- Timing: the four smaller collections (~55k points) land in **minutes**;
  `publications` (430k points) is limited by the embedded-format reader and
  takes **a few hours** on a laptop — run it in `tmux`/`screen` or overnight.
  Practical order: grab the small ones first, then let publications grind:
  ```bash
  python load_all.py --url ... --collections copernicus_docs marine_docs cds_docs eqc_qa
  python load_all.py --url ... --collections publications   # long — tmux it
  ```
- **Resumable / idempotent**: re-running skips collections whose point count
  already matches; `--recreate` forces a clean re-copy. An interrupted
  collection is re-copied from scratch on the next run.
- Remote/managed cluster: `--url https://<cluster>.cloud.qdrant.io --api-key <key>`
- Already downloaded the tarballs? Untar each `indexes/qdrant_<name>.tar.gz`
  into `<dir>/qdrant_<name>/` and pass `--source-dir <dir>`.

The script verifies every collection (`server count == source count`) and
aborts loudly on mismatch. Expected final state:

```
copernicus_docs 1418 · marine_docs 29249 · cds_docs 23341 · eqc_qa 1274 · publications 430066
```

## 5 · Query it

### Keyword / BM25 (no embedding model needed)

```python
from qdrant_client import QdrantClient, models
from fastembed import SparseTextEmbedding

c = QdrantClient(url="http://localhost:6333")
bm25 = SparseTextEmbedding(model_name="Qdrant/bm25")

def sparse(q):
    r = list(bm25.embed([q]))[0]
    return models.SparseVector(indices=r.indices.tolist(), values=r.values.tolist())

hits = c.query_points("publications", query=sparse("marine heatwave detection SST"),
                      using="sparse", limit=5, with_payload=True)
for h in hits.points:
    print(round(h.score, 2), h.payload["title"][:70])
```

### Dense & hybrid (needs query embeddings)

The corpus vectors are `gemini-embedding-2-preview`, `RETRIEVAL_QUERY` task,
768-d, **L2-normalized** — embed queries the same way:

```python
import numpy as np
from google import genai
from google.genai import types

g = genai.Client(api_key=GEMINI_KEY)
def dense(q):
    r = g.models.embed_content(model="gemini-embedding-2-preview", contents=q,
        config=types.EmbedContentConfig(task_type="RETRIEVAL_QUERY", output_dimensionality=768))
    v = np.array(list(r.embeddings[0].values), dtype=np.float32)
    return (v / np.linalg.norm(v)).tolist()

# hybrid: dense + BM25 fused with RRF, server-side
hits = c.query_points(
    "publications",
    prefetch=[
        models.Prefetch(query=dense(q), using="dense", limit=20),
        models.Prefetch(query=sparse(q), using="sparse", limit=20),
    ],
    query=models.FusionQuery(fusion=models.Fusion.RRF),
    limit=5, with_payload=True)
```

No Gemini access? Two options: BM25-only (above, surprisingly strong on this
corpus), or re-embed the chunks with an open model — `REBUILD.md` §A is the
recipe (`chunks/*.jsonl` carry the raw text; swap is ~20 lines).

### Filters (payload indexes are in place)

```python
# everything linked to ERA5, full-text chunks only
flt = models.Filter(must=[
    models.FieldCondition(key="linked_products",
                          match=models.MatchValue(value="reanalysis-era5-single-levels")),
    models.FieldCondition(key="chunk_type", match=models.MatchValue(value="text")),
])
c.query_points("publications", query=sparse("wind energy assessment"),
               using="sparse", query_filter=flt, limit=5)

# QUID (quality) docs for one CMEMS product
flt = models.Filter(must=[
    models.FieldCondition(key="product_id",
                          match=models.MatchValue(value="GLOBAL_MULTIYEAR_PHY_001_030")),
    models.FieldCondition(key="doc_type", match=models.MatchValue(value="QUID")),
])
c.query_points("marine_docs", query=sparse("assimilated observations"),
               using="sparse", query_filter=flt, limit=5)
```

## 6 · Ops notes

- **Backups**: `curl -X POST http://localhost:6333/collections/publications/snapshots`
  (or just stop the container and copy `qdrant_storage/`).
- **Upgrades**: bump the image tag in `docker-compose.yml`; storage is
  forward-compatible across minor versions.
- **Memory**: `on_disk_payload` is enabled; vectors stay in RAM
  (~1.5 GB total). For tighter RAM add on-disk HNSW/quantization —
  see Qdrant docs.
- The MCP server (`scripts/marine_rag/rag_server.py`) is **server-first**: if a
  Qdrant server is reachable at `QDRANT_URL` (default `http://localhost:6333`)
  and carries the collection, it uses it — publications queries drop from
  minutes (embedded) to ~30 ms. It silently falls back to its embedded copies
  when the server is down; set `QDRANT_URL=""` to force embedded mode.

## 7 · Troubleshooting

| symptom | fix |
|---|---|
| `401` on download | token lacks read access, or you're not a collaborator on the dataset |
| `load_all.py` can't reach server | `docker compose ps`, `curl localhost:6333/readyz`; port 6333 busy → change mapping |
| count MISMATCH abort | re-run (it re-copies the failed collection); check server disk space |
| slow upserts | expected on spinning disks; use `--collections` to prioritize what you need first |
| Apple Silicon | works out of the box (multi-arch image) |