Retrieve the Most Similar Prior SARs by Narrative Meaning

Objective

SAR-narrative quality is analyst-dependent: the same fact pattern yields a strong narrative from a senior and an inconsistent one from a junior, and the examiner sees the variance. This recipe retrieves the most semantically similar prior SARs for a new case, so the analyst starts from proven language for the matching typology — consistency by construction.

Step 1: Your SAR history with narrative embeddings

DROP TABLE IF EXISTS sar_hist;

CREATE TABLE sar_hist (
    sar_id    TEXT PRIMARY KEY,
    typology  TEXT,
    narrative TEXT,
    emb       VECTOR(384)
);

INSERT INTO sar_hist VALUES
 ('SAR-9001','structuring',
  'Customer made repeated sub-10k cash deposits across several days totaling well over the reporting threshold.',
  EMBED('Customer made repeated sub-10k cash deposits across several days totaling well over the reporting threshold.')),
 ('SAR-9002','layering',
  'Funds moved rapidly through three shell accounts before an offshore cashout.',
  EMBED('Funds moved rapidly through three shell accounts before an offshore cashout.'));

Step 2: Find the closest prior SARs for a new case

SELECT sar_id,
       typology,
       COSINE_SIMILARITY(
         emb,
         EMBED('Client broke a large cash amount into many deposits under 10000 to avoid the CTR threshold.')
       ) AS sim
FROM   sar_hist
ORDER  BY sim DESC
LIMIT  2;

Result — the new structuring case correctly retrieves the prior structuring SAR ahead of the layering one:

Why it matters

The analyst drafts from a proven, same-typology narrative instead of a blank page — faster and more consistent for the regulator. This is the retrieval half of a grounded SAR-drafting workflow; pair it with an agentic reviewer to draft, and a tamper-evident ledger to prove the process.