Deterministic baseline
Query-token coverage: matched query tokens divided by all query tokens. Simple, inspectable, sensitive to wording and broad terms.
Executed in this demoAI ENGINEER - INTERN / WORKING PROOF OF CONCEPT
Before a keyword
becomes a claim.
A small decision bench for marketplace sellers: match search intent to catalog facts, inspect the misses, then draft a listing without inventing product features.
Synthetic catalogLocal computationNo Amazon connection
RETRIEVAL, WITH A BRAKE
Which product actually fits?
Try a failure:
Scores are lexical similarity, not confidence or conversion probability. Thresholds are exploratory, not calibrated. Both methods see exactly the same catalog and queries.
Important: a high score can still retrieve the wrong intent. “Bottle cleaning brush” shares the word “bottle” but is not the bottle itself. Nothing here automatically creates negative keywords.
FROM CATALOG FACT TO COPY
Keep every draft reviewable.
This is a token allowlist warning, not semantic verification or an Amazon policy checker. It can flag safe synonyms and miss a misleading combination of known words. The 180-character budget is this prototype's choice, not a marketplace policy claim.
SAME FIXTURE, VISIBLE FAILURES
Does the method earn the next step?
36 manually authored synthetic queries: 24 literal, 4 unrelated, 4 paraphrases and 4 accessory-intent traps. Development fixture, not a held-out benchmark. IDF is fitted only on the six catalog descriptions.
Precision = correct accepted / all accepted. Recall = correct accepted / 28 labeled product queries. “Rejected” can mean a safe abstention or a missed relevant query. Change thresholds in Query fit to rerun both methods.
Potential impact: help a seller review keyword-to-product fit and unsupported listing language before acting. No sales, advertising-spend or time-saved improvement has been measured.
Watch the narrated walkthrough (1:12)
Actual interactive browser recording. Synthetic catalog, local lexical models and explicit failure cases.
IMPLEMENTATION / FEASIBILITY NOTE
A baseline first. A model only when earned.
TF-IDF retrieval
Smoothed IDF and L2 cosine over catalog-only vocabulary. Downweights shared terms; out-of-vocabulary words are invisible to the vector and can inflate a match.
Executed in this demoDense retrieval + reranking
A candidate next experiment for synonyms and intent. Needs a seller-labeled, untouched test set and measured latency/cost. Could still confuse products with accessories.
Proposed, not executedArchitecture, limitations and source notes
Flow: local synthetic catalog → tokenizer/IDF index → query ranking → threshold + gap gate → human review. The separate title guard compares tokens with the selected product's facts. No backend, model API, cookies or seller-data upload.
Python reference: the source package includes a reproducible Python implementation, tests and evaluation results. The deployed app runs the JavaScript engine in the browser. This is not an existing eComSuite integration.
Next validation: obtain seller permission and product facts, collect independent query labels, split by product/query family before tuning, test Hindi/English and accessory ambiguity, compare dense retrieval on the same untouched set, then measure latency and human-review burden. Never automate bid/listing edits from these lexical scores.
Research checked 14 Nov 2026. Company logo, purple palette and Urbanist typography observed on its site. The layout and workflow are an independent concept, not a copy of its authenticated product.