# Session: stemwijzer — Parliamentary Embedding Pipeline
Updated: 2026-03-22T16:00:00Z

## Goal
2D political compass + motion similarity search from parliamentary votes + motion text.
Full historical coverage 2016–2026, precomputed similarity cache, fused (SVD + text) embeddings.

## Constraints
- DuckDB only (`data/motions.db`); open/close `duckdb.connect(self.db_path)` per method
- Vectors stored as JSON text (no external vector DB)
- Logging via `logging.getLogger(__name__)`; no `print()` in library modules
- Tests run offline (network monkeypatched) — use `.venv/bin/python -m pytest -q`
- Do NOT modify `app.py` or `scheduler.py`
- Use `.venv/bin/python` (Arch Linux system Python is externally managed)

## Current DB State (verified 2026-03-22 ~16:00)

| Table | Rows |
|---|---|
| motions | 10,613 |
| embeddings | 10,753 |
| svd_vectors | 24,528 |
| fused_embeddings | **10,613** (1:1 with motions, 0 duplicates) |
| similarity_cache | **212,206** (top_k=20, all annual windows) |
| mp_votes | 199,967 |
| mp_metadata | 798 |

## Annual Window Coverage

| Year | Motions | Fused | Similarity |
|---|---|---|---|
| 2016 | 132 | 132 | 2,640 |
| 2017 | 30 | 30 | 600 |
| 2018 | 100 | 100 | 2,000 |
| 2019 | 3 | 3 | 6 |
| 2020 | 0 | 0 | 0 (no data) |
| 2021 | 0 | 0 | 0 (no data) |
| 2022 | 4,116 | 4,116 | 82,320 |
| 2023 | 621 | 621 | 12,420 |
| 2024 | 948 | 948 | 18,960 |
| 2025 | 3,715 | 3,715 | 74,300 |
| 2026 | 948 | 948 | 18,960 |

## Completed This Session
- [x] Text embeddings: ran with real OpenRouter API at batch_size=200 → 10,753 embedding rows
- [x] Re-ran `extract_mp_votes` on all motions → 111,978 new rows (party-level votes backfilled)
- [x] SVD re-run (annual 2016–2026) with full vote data → 24,528 svd_vector rows
- [x] Fixed `store_fused_embedding` double-counting bug: added DELETE before INSERT
- [x] Cleaned and re-ran fusion → 10,613 fused rows, zero duplicates
- [x] Re-ran similarity cache top_k=20 for all 9 active windows → 212,206 rows
- [x] Test suite: **34 passed, 2 skipped** ✅

## Key Decisions
- `store_fused_embedding` (database.py line 686): Now does DELETE+INSERT instead of plain INSERT to prevent duplicates on re-runs.
- Annual windows chosen for historical political compass (2016–2026).
- top_k=20 for similarity cache.
- Party-level votes (e.g. `{"PVV": "voor"}`) handled in `extract_mp_votes` — actor without comma → `party=actor_name`.

## Open Items (not blocking, data coverage gaps)
1. **2020–2021 data gap**: No motions in DB at all. Need to run downloader with `--start-date 2019-01-01 --end-date 2021-12-31` if data exists in API.
2. **2024 gap ~3,020 motions**: OData API has ~3,968 2024 motions, only 948 in DB. Root cause unclear — needs investigation of URL-based dedup in `insert_motion`.
3. **"Verworpen." dedup**: Short-text motions (title="Verworpen.") get spurious similarity=1.0. UI/query layer should filter `score < 0.999 OR title != 'Verworpen.'`.
4. **svd_vectors has duplicates**: 2025 has 7,430 rows for 3,715 motions (2x). Doesn't affect fused_embeddings (DELETE+INSERT handles it) but wastes space. Low priority.

## Key File Paths
- DB: `data/motions.db`
- Venv: `.venv/bin/python`
- Pipeline entry: `pipeline/run_pipeline.py`
- Fusion: `pipeline/fusion.py`
- SVD: `pipeline/svd_pipeline.py`
- Text embeddings: `pipeline/text_pipeline.py`
- MP votes extraction: `pipeline/extract_mp_votes.py`
- Database layer: `database.py`
- Similarity compute: `similarity/compute.py`
- Similarity lookup: `similarity/lookup.py`
- Tests: `tests/` (pytest, offline)

## Branch
`main`