How retrieval works

Recall: deterministic Okapi BM25

k1=1.5, b=0.75, frozen stopwords, stable tie-breaks. The same corpus and query produce the same ranking, forever.

The scoring function

For query terms q, document d with term frequency f(t,d), length |d| and corpus average length avgdl:

score(q,d) = SUM over t in q of
    IDF(t) * ( f(t,d) * (k1 + 1) )
           / ( f(t,d) + k1 * (1 - b + b * |d| / avgdl) )

IDF(t) = ln( 1 + (N - df(t) + 0.5) / (df(t) + 0.5) )
k1 = 1.5    b = 0.75    N = live documents in the space

Determinism guarantees

• Terms are lowercased and stopword-filtered by a frozen list that is part of the contract.
• Documents with a positive score only are returned; zero-score documents never pad a result.
• Ordering is score descending, then created_at ascending, then memory_id ascending.
• No randomness, no sampling, no approximate index. The same rows give the same bytes.
• Every response states ranking: "Okapi BM25 (k1=1.5, b=0.75)" so a caller can reproduce it.

Filters

space is a hard partition, not a ranking hint. tags are conjunctive: a memory must carry *every* requested tag. include_quarantined lists quarantined rows (id, status, reason) with their text withheld; only GET /aether33/api/memories/{id} with the X-CAIN-Reviewer secret shows the text.

There is no cross-space ranking and no cross-tenant search. Those are not options that default to off; they are not implemented.

What BM25 cannot do

It will not find a paraphrase with no shared vocabulary. “car” does not retrieve “automobile”. If you need that, put a vector index in front of AETHER-33 and keep this service as the governed record of what was actually written and who said it — do not pretend a cosine similarity is provenance.