Shopping Copilot
Built for TikTok TechJam 2026 · Track 4

Shopping Copilot

Find the purchased product earlier and rank it higher — over 50,000 catalog items, in at most 10 turns.

—
TechnicalScore
HitRate@10 — MRR — MTTC —
Inspect verified results
Competition Evidence Official public evaluator · — sessions
Offline CPU · Python standard library · No API keys
Private 800-session performance remains unknown.
Competition Data Contract
What we agreed to solve
Competition Evidence

SCENARIO MIX

Catalog Samples
Text-only metadata — no product images
Step 1 of 6
Scenario Replay
Real public sessions from the official evaluator
Competition Evidence

Route

Hard Constraints

Soft Preferences

Negative Constraints

Ask Attribute

Top-10 Recommendations
Turn 1
Step 2 of 6
Mechanism Inspection
How the pipeline processes a real request
Competition Evidence

Pipeline

Message → Intent Router → Versioned Constraint State → SQLite FTS5 Retrieval → Rule Reranking → Candidate-driven Clarification → Top-10 parent_asin
Experiments We Did Not Ship
Negative Result
Step 3 of 6
Evaluation Evidence
Official public-set evaluator result · Not hidden-set evidence
Competition Evidence

VERSION COMPARISON

Version HitRate@10 MRR MTTC TechnicalScore

PER-SCENARIO BREAKDOWN (V1.3)

Scenario N HitRate@10 MRR MTTC

REPRODUCE


        

          
        

      

    


    

      
      Step 4 of 6
      
    

  


  
  
  
  

    

      

        

          
Beyond the Benchmark — Transparent Ads

          
Demo-only commercial extension

        

        Demo Only
      


      

        Demo-only commercial extension.

        Never called by the official evaluator. Does not alter organic Top-10 ranking.
      


      
PRESET EXPERIMENT

      

        

          
Campaign A

          
Bid$1.00

          
Relevance0.82

          
eCPM—

        

        

          
Campaign B

          
Bid$5.00

          
Relevance0.12

          
eCPM—

        

      


      

        
      


      

      

        This chapter proves: transparent ad auction · relevance-aware monetization · budget accounting · official/demo path isolation · impact potential.

        Current scope: impression auction and budget accounting. No click/conversion path, CTR, or GMV.
      

    


    

      
      Step 5 of 6
      
    

  


  
  
  
  

    

      

        
Deliverables & Limitations

        Competition Evidence
      


      

        
          📂 Public Repository
        
        
          📄 Technical Report
        
        
          🔁 Reproduction Instructions
        
        
          🎥 Demo Video (placeholder)
        
      


      

        
Limitations

        
    
          
  • Public-set gains fix generic error classes but do not substitute for the hidden 800-session validation.
    • 
          
  • Popularity tiebreaker band (5.0) is tuned on public 200; a conservative band=3 (+0.03) is available if hidden set regresses.
    • 
          
  • The optional Qwen layer improves vague-intent classification but is not required for reported scores.
    • 
          
  • No dense/vector recall — proven unnecessary since recall is 100% saturated (200/200).
    • 
          
  • Ad engine is demo-only with simulated inventory and budgets.
    • 
        

      


      

        What we ship: a deterministic, offline, reproducible scored Agent.

        What we demonstrate beyond the score: transparent commercial extensions.

        What remains unknown: organizer-private 800-session performance.
      


      

      

    


    

      
      Step 6 of 6 — Tour Complete
      Explore all evidence →