LLM Lifecycle
Like other software products, LLM models have a life and eventual deprecation cycle. The following table describes the Aisera Gen AI platform features that correspond to this cycle.
Data Collection and Curation Gathering and cleaning the data used to train or fine-tune the model.
he Aisera Gen AI platform includes data mapping and integration jobs that omit non-mapped data fields, except where customized.
Training and Fine-Tuning Building a base model or customizing an existing LLM to fit specific tasks.
The Aisera Gen AI platform allows users and implementation engineers to choose a specific domain when creating a bot. Each domain includes a default Ontology that provides agentic capability for specific tasks.
Evaluation Testing model performance using metrics like accuracy, toxicity, bias, and robustness.
The Request Analyzer, Conv AI 2.0, AI Lens, and Test Suite features allow customers to test model performance. We also have the following Benchmark: Aisera LLM Benchmarking
Deployment
Infrastructure Setup Choosing where and how the model runs (cloud, on-premise, edge)
The Aisera Admin and Customer Teams deploy the cloud or on-prem administrative setup for the customer, based on their needs.
Serving Making the model available via APIs or integration into Apps
The Aisera Gen AI platform makes the models available via integration into applications/bots. Aisera admins can switch the bots used for a specific tenant. Integration for data transfer is available via Ingestion APIs.
Scalability Ensuring performance at different usage volumes
The Aisera Gen AI platform uses elasticity via AWS
Monitoring and Maintenance
Performance Monitoring Tracking usage, latency, and response quality
Feedback Loops Collecting human feedback or interaction data for future improvements.
Error and Drift Detection Identifying issues such as model hallucinations or concept drift over time.
Compliance and Governance
Security and Privacy Managing access, protecting data, and complying with regulations (such as GDPR)
Bias and Fairness Auditing Regular checks to ensure responsible AI use.
Documentation and Traceability Keeping detailed records for reproducibility and accountability.
Iteration and Updates
Retraining or Fine Tuning Updating the model based on new data or feedback.
Versioning Managing multiple model versions and rolling out changes safely.
So far, the Aisera GPT bots have two versions, 1.0 and 2.0. The differences are described here:
Retirement
Decommissioning Phasing out outdated models.
Archiving Storing model artifacts and metadata for compliance or future reference.
Before you clone or archive your bot, use the icons at the top of your Data Source Details window to Clone your bot, retire Knowledge Documents, or clear the Tickets Index.
Last updated
Was this helpful?