LLM Agent / Map Data Quality
A small LLM-agent style workflow for POI data quality judgment using mock map, merchant, user feedback, and spatial evidence.
Evidence chunks
10+
mock multi-source signals
Top-K retrieval
5
ranked evidence
Workflow steps
7
visible tool trace
overview
The FastAPI backend loads mock POI cases, routes the task type, retrieves evidence chunks with deterministic top-k scoring, compares multi-source signals, computes spatial consistency, makes a conservative quality decision, and uses OpenRouter/Gemini only to explain the fixed decision.
role
Portfolio demo: built a map-data quality agent console for address correction, merchant status judgment, duplicate POI detection, evidence retrieval, tool tracing, and explainable manual-review decisions.
backend runner
Calls /api/poi/run on the FastAPI wrapper at https://ruizelab-api.onrender.com.
POI scenario
Request payload
case_relocation_luckin · top_k 5 · OpenRouter/Gemini explanation enabled
backend contract
Start the backend with uvicorn backend.main:app --reload --port 8000. The UI displays mode and provenance so recorded artifacts are clearly distinguished from live computation.
Configure the inputs and run the backend to display the original project trace and outputs here.
architecture flow
The live pipeline trace appears in the backend runner after execution. This section shows the original project components that the backend wraps.
planner
Classifies a POI query as address correction, duplicate detection, or merchant status checking.
retrieval
Ranks mock map, merchant, user feedback, historical, order, photo, and external listing evidence with top-k scoring.
spatial
Calculates haversine distance or pair distance to judge relocation and duplicate POI consistency.
verifier
Groups evidence by support label, highlights conflicts, and keeps risky POI updates in manual review.
explainer
Uses OpenRouter/Gemini for a concise Chinese explanation while preserving the deterministic decision.
tools and models
planner
Classifies a POI query as address correction, duplicate detection, or merchant status checking.
retrieval
Ranks mock map, merchant, user feedback, historical, order, photo, and external listing evidence with top-k scoring.
spatial
Calculates haversine distance or pair distance to judge relocation and duplicate POI consistency.
verifier
Groups evidence by support label, highlights conflicts, and keeps risky POI updates in manual review.
explainer
Uses OpenRouter/Gemini for a concise Chinese explanation while preserving the deterministic decision.
example input
判断瑞幸咖啡中关村大街店是否地址正确，是否疑似搬迁
final result
The backend returns original POI records, scored top-k evidence, a seven-step tool trace, grouped evidence comparison, a deterministic quality decision, an optional Gemini explanation, limitations, and provenance.
limitations