code/scripts/03_retrieve.py

"""Step C. GPU retrieval for one channel and one query shard.



    CUDA_VISIBLE_DEVICES=0 nice -n 19 ionice -c3 python scripts/03_retrieve.py \

        --split test --channel e5_small --k 100 --shard 0 --nshards 2



Two GPUs run the same command with --shard 0 and --shard 1. They share nothing, so either

shard can be re-run alone after a failure without redoing the other.



The channel matrices are expected at BERX_EMB/<channel>/<split>_{s1,pool}.npy, fp16, L2

normalised, in source row order. `--verify` checks the shapes against ids.json before

spending an hour on a matrix that turns out to be the wrong model - the two 1024-d channels

are interchangeable by shape and not by content, so shape alone will not catch a mix-up.

"""
from __future__ import annotations

import argparse
import json
import os
import sys
import time

import numpy as np
import pandas as pd

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from berx import paths as P  # noqa: E402
from berx import retrieve as R  # noqa: E402


def verify_ids(channel: str, split: str, which: str, prep: str, parquet: str, n_check=200_000):
    """Check the embedding's ids.txt against the prepared source, row by row.



    Only the first and last `n_check` rows are compared in full: reading 10M ids takes longer

    than the check is worth, and any offset introduced mid-file shifts every subsequent row,

    so the tail catches it.

    """
    ids_path = P.emb_ids(channel, split, which)
    if not os.path.exists(ids_path):
        print(f"  ! no ids.txt for {channel}/{split}_{which}; row order UNVERIFIED")
        return
    want = pd.read_parquet(os.path.join(prep, parquet), columns=["entity_id"])["entity_id"]
    with open(ids_path) as fh:
        got = [ln.strip() for ln in fh]
    assert len(got) == len(want), \
        f"{ids_path}: {len(got):,} ids but the prepared {parquet} has {len(want):,} rows"
    k = min(n_check, len(want))
    head_bad = [i for i in range(k) if got[i] != want.iloc[i]]
    tail_bad = [i for i in range(len(want) - k, len(want)) if got[i] != want.iloc[i]]
    assert not head_bad and not tail_bad, (
        f"{ids_path}: row order does not match {parquet}. "
        f"First mismatch at row {(head_bad or tail_bad)[0]:,}. "
        f"Every candidate index from there on would be wrong.")
    print(f"  ids verified: {channel}/{split}_{which}  {len(got):,} rows, "
          f"head+tail {k:,} each match")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--split", default="test")
    ap.add_argument("--channel", required=True,
                    help="a configured channel, or an ad-hoc one built by 08_embed.py "
                         "(nameonly / addronly)")
    ap.add_argument("--k", type=int, default=100)
    ap.add_argument("--shard", type=int, default=0)
    ap.add_argument("--nshards", type=int, default=1)
    ap.add_argument("--device", default="cuda:0")
    ap.add_argument("--qchunk", type=int, default=4096)
    ap.add_argument("--pblock", type=int, default=262_144)
    ap.add_argument("--resident", default="auto", choices=["auto", "yes", "no"])
    ap.add_argument("--max-queries", type=int, default=0,
                    help="debug: cap queries per country. Verifies the whole path on a busy "
                         "box in seconds instead of hours. Never use for a real run.")
    ap.add_argument("--verify-only", action="store_true",
                    help="check shapes, normalisation and row order, then exit")
    a = ap.parse_args()

    prep = os.path.dirname(P.work("prep", a.split, "_"))
    meta = json.load(open(os.path.join(prep, "ids.json")))
    qc = np.load(os.path.join(prep, "country_s1.npy"))
    pc = np.load(os.path.join(prep, "country_pool.npy"))

    qp = P.emb(a.channel, a.split, "s1")
    pp = P.emb(a.channel, a.split, "pool")
    for p in (qp, pp):
        assert os.path.exists(p), f"missing embedding matrix: {p}"
    Q, Pm = np.load(qp, mmap_mode="r"), np.load(pp, mmap_mode="r")
    # channel_cfg, not CHANNELS: the rescue channels (nameonly, addronly) are built on demand
    # by 08_embed.py and have no declared entry.
    want_dim = P.channel_cfg(a.channel)["dim"]
    assert Q.shape == (meta["n_s1"], want_dim), \
        f"{qp} is {Q.shape}, expected ({meta['n_s1']}, {want_dim})"
    assert Pm.shape[1] == want_dim, f"{pp} is {Pm.shape[1]}-d, expected {want_dim}"

    # A SUBSET pool (F4 covers only the address-less records) carries rows.npy mapping its
    # rows back to rows of the FULL pool. Everything downstream indexes the full pool, so
    # without this remap candidate index i means row i of a 265k subset while the fuse step
    # reads it as row i of a 10M pool - wrong, and with nothing to signal it.
    rows_path = P.emb_rows(a.channel, a.split, "pool")
    prow = np.load(rows_path) if os.path.exists(rows_path) else None
    if prow is not None and len(prow) != meta["n_pool"]:
        assert len(prow) == Pm.shape[0], \
            f"rows.npy has {len(prow):,} entries but the matrix has {Pm.shape[0]:,} rows"
        print(f"  subset pool: {Pm.shape[0]:,} of {meta['n_pool']:,} rows; "
              f"candidate indices remapped through rows.npy")
        pc = pc[prow]
    else:
        prow = None
        assert Pm.shape[0] == meta["n_pool"], \
            f"{pp} has {Pm.shape[0]:,} rows, expected {meta['n_pool']:,}"
    nrm = np.linalg.norm(np.asarray(Q[:256], dtype=np.float32), axis=1)
    assert abs(float(nrm.mean()) - 1.0) < 0.02, \
        f"query rows are not L2-normalised (mean norm {nrm.mean():.4f}); cosine would be wrong"

    # Verify row order by ENTITY ID, not by row count. A matrix with the right number of rows
    # in the wrong order produces candidate indices that are wrong for every row after the
    # first disturbance, and nothing downstream - not the validator, not the F0.5 - reports
    # anything other than a mediocre score.
    verify_ids(a.channel, a.split, "s1", prep, "source1.parquet")
    if prow is None:
        verify_ids(a.channel, a.split, "pool", prep, "pool.parquet")

    if a.verify_only:
        print("verify-only: shapes, normalisation and row order all check out.")
        return

    print(f"channel {a.channel}  k={a.k}  shard {a.shard}/{a.nshards}  {a.device}")
    t0 = time.time()
    res = R.run_channel(qp, pp, qc, pc, a.k, a.device, a.shard, a.nshards,
                        qchunk=a.qchunk, pblock=a.pblock, resident=a.resident,
                        max_queries=a.max_queries or None)
    if prow is not None:
        res["c"] = prow[res["c"]].astype(np.int32)      # subset row -> full pool row
    df = pd.DataFrame(res)
    out = P.work("retr", f"{a.split}_{a.channel}_s{a.shard}of{a.nshards}.parquet")
    df.to_parquet(out, index=False)
    el = time.time() - t0
    print(f"wrote {out}  {len(df):,} pairs  {el/60:.1f} min  "
          f"({len(df)/max(el,1e-9):,.0f} pairs/s)")


if __name__ == "__main__":
    main()