Fast Answer offline architecture
Historical chat logs and production feedback feed separate add and update-delete Catalog paths that converge at human review. A second pipeline trains the Decision model. Their versioned outputs converge into an immutable release bundle stored in object storage.
Catalog Construction
INPUT SIGNALS
Historical Chat Logs
Production Feedback
SPARK
Daily Aggregates
ROLLING WINDOW
30-Day Frequency
ADD PATH
Candidate Filtering
SPARK
Update / Delete Candidates
WORKFLOW
Human Review
Add · Update · Delete
VERSIONED
Approved
Catalog
Decision Model Training
OBJECT STORAGE
Experiment Logs
SPARK
Feature Generation
TRAINING JOB
T-Learner Training
HOLDOUT DATA
Off-Policy Evaluation
VERSIONED OUTPUT
Decision Model
+ Decision Threshold
VALIDATE + PACKAGE
Release Build
FAST ANSWER RELEASE
Immutable Bundle
Catalog · Model · Threshold
PUBLISH
Versioned Object Storage