ShopMate AI: A Grounded Agentic Shopping Assistant with Claim Verification, Three-Valued Requirement Matching, Bounded Replanning, and a Consent-Gated Mock Checkout

An LLM-based shopping assistant can search, compare and recommend products; however, it falls down in three typical ways: it claims things about the product that the data doesn't support, it takes silence for pass or fail, and it acts on behalf of the user without a traceable approval. We present ShopMate AI, an agentic shopping assistant for the Indian market, based on the LangGraph pipeline, where the language model plans and explains but never decides truth. The product facts are extracted from listing specifications into typed values, and a fact exists only if the listing explicitly says it; contradictory listings yield no fact. The requirements of the user are matched using deterministic rules to five outcomes (met, not met, unknown, conflict, and unsupported) such that silence means "could not confirm" and not guessing. Each sentence of explanation is drafted as a claim and validated deterministically by the verifier using the data retrieved, allowing one rewording and removing unsupported claims. If there is no match for the request, then a bounded replanning step provides code-built relaxation suggestions (budget increase or removal of one requirement, but not a specific brand). Purchase requests are constrained by a server-side cart mandate which is signed with HMAC-SHA256 following explicit consent and then checked during mock checkout. On a snapshot of 480 Indian e-commerce listings in five product categories and 66 evaluation queries, a fault-injection test shows that the verifier rejects 99.9% of 12,132 intentionally corrupted claims and none of the 7,321 true claims. Live language-model runs across the 66 queries (99 runs) parse every category and budget correctly (100%), and none of the 1,350 drafted claims failed verification, while a no-tools baseline asked to recommend products names items that match a current listing in only 6 of 180 cases (3.3%). Offline, replanning succeeds in finding products for 5 out of 6 core queries that produce nothing on the first pass. Requirement matching rules have not been validated with human labelling.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24361156
Primary Topic
Information Retrieval and Search Behavior
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

ShopMate AI: A Grounded Agentic Shopping Assistant with Claim Verification, Three-Valued Requirement Matching, Bounded Replanning, and a Consent-Gated Mock Checkout

Yisol Choi, Suresh Kandasamy, Shreya H Rao, Ritika Reddy Alla et al.
Zenodo (CERN European Organization for Nuclear Research)
Information Retrieval and Search Behavior
preprint

ShopMate AI: A Grounded Agentic Shopping Assistant with Claim Verification, Three-Valued Requirement Matching, Bounded Replanning, and a Consent-Gated Mock Checkout

Yisol Choi, Suresh Kandasamy, Shreya H Rao, Ritika Reddy Alla, Ankith D Prabhu
preprint en

Abstract

An LLM-based shopping assistant can search, compare and recommend products; however, it falls down in three typical ways: it claims things about the product that the data doesn't support, it takes silence for pass or fail, and it acts on behalf of the user without a traceable approval. We present ShopMate AI, an agentic shopping assistant for the Indian market, based on the LangGraph pipeline, where the language model plans and explains but never decides truth. The product facts are extracted from listing specifications into typed values, and a fact exists only if the listing explicitly says it; contradictory listings yield no fact. The requirements of the user are matched using deterministic rules to five outcomes (met, not met, unknown, conflict, and unsupported) such that silence means "could not confirm" and not guessing. Each sentence of explanation is drafted as a claim and validated deterministically by the verifier using the data retrieved, allowing one rewording and removing unsupported claims. If there is no match for the request, then a bounded replanning step provides code-built relaxation suggestions (budget increase or removal of one requirement, but not a specific brand). Purchase requests are constrained by a server-side cart mandate which is signed with HMAC-SHA256 following explicit consent and then checked during mock checkout. On a snapshot of 480 Indian e-commerce listings in five product categories and 66 evaluation queries, a fault-injection test shows that the verifier rejects 99.9% of 12,132 intentionally corrupted claims and none of the 7,321 true claims. Live language-model runs across the 66 queries (99 runs) parse every category and budget correctly (100%), and none of the 1,350 drafted claims failed verification, while a no-tools baseline asked to recommend products names items that match a current listing in only 6 of 180 cases (3.3%). Offline, replanning succeeds in finding products for 5 out of 6 core queries that produce nothing on the first pass. Requirement matching rules have not been validated with human labelling.

Zenodo (CERN European Organization for Nuclear Research)
Dayananda Sagar University (IN)
Information Retrieval and Search Behavior
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon