Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads) v1.3
This page introduces the core entities you use to build Gen AI applications in AI Factory and how they relate. It orients you before you dive into how‑tos or SDK details.
Assistant (orchestrator)
An Assistant is the orchestrator of your application logic:
- Executes the directive (goal) and coordinates retrieval and tool use.
- Attaches to Knowledge (for RAG) and uses Tools (for actions, APIs).
- Runs with a model endpoint configured by your platform (see Model Serving).
Learn more: Assistants
Ruleset (policy and constraints)
Rulesets constrain behavior and guide output:
- Capture policy: tone, style, safety guidelines, guardrails.
- Apply to an Assistant to enforce boundaries across runs.
Learn more: Rulesets
Knowledge (retrieval)
Your knowledge stack grounds the model on enterprise data:
- Knowledge Bases: embedded content from tables/files with metadata for filters.
- Retrievers: define search parameters (top‑K, thresholds, filters) for hybrid search.
- Data Lake: repository for assets and files that feed knowledge ingestion.
Learn more:
Tools (actions and integrations)
Tools allow Assistants to call external systems and internal APIs:
- Wrap REST endpoints or internal services with a typed interface.
- Provide authentication and input/output mapping.
- Govern which tools an Assistant can call via Rulesets and configuration.
Learn more: Tools
Structures (workflows and building blocks)
Structures are higher‑level building blocks (tasks, pipelines, workflows) you can compose:
- Encapsulate reusable logic patterns.
- Integrate with Assistants or call Tools and retrieval internally.
Learn more: Structures
Threads (state and observability)
Threads persist conversation state and provide auditability:
- Record prompts, tool calls, retrieved context, and responses.
- Enable review, debugging, and compliance.
Learn more: Threads
Models and endpoints (generation)
Assistants use model endpoints for generation:
- Use Model Serving for private inference.
- Call endpoints internally (cluster‑local) or externally (portal + access key).
- Use operation‑specific paths: chat (
/v1/chat/completions), embeddings (/v1/embeddings), rerank (/v1/ranking).
Learn more: Access KServe endpoints
Putting it together
At a high level:
1) Prepare Knowledge (knowledge bases with hybrid search) and Tools. 2) Configure an Assistant with a directive, attach Knowledge and Tools, and select a model endpoint. 3) Use Threads to observe behavior, debug traces, and govern interactions.
Build guides: