Technical Glossary
This glossary defines standard terminology, third-party libraries, and internal abstractions used across the AI Dashboard codebase.
Agentic Pipeline
A multi-step reasoning and data processing workflow where autonomous or semi-autonomous software entities (agents) execute distinct responsibilities sequentially or conditionally.
Composed Dashboard Spec
The finalized JSON contract produced by the Layout Agent. It encapsulates dashboard metadata, unified styling configs, individual chart specs, pre-fetched raw data arrays, and SQL tracing queries. Validated via Pydantic on the backend and mapped to strict TypeScript interfaces on the frontend.
Data Agent
The second stage of the generation pipeline. It receives ChartGoal structures, formulates database-specific SQL queries based on active schemas, runs proactive security sanitization checks, executes read-only operations via SQLAlchemy, and packages raw row records.
Intent Classification
A routing mechanism invoked during dashboard refinement (POST /dashboard/refine). It analyzes the user's conversational feedback against current chart parameters to categorize the intended action (e.g., chart_type_change, title_change, add_chart, sql_modification) or flag when clarification is needed.
LangGraph
An orchestration extension for LangChain designed to build robust, stateful, multi-actor applications with cycles. The AI Dashboard relies on LangGraph (StateGraph) to govern transitions between agent stages and stream partial execution events.
Layout Agent
The fourth and final stage of the generation pipeline. It processes an array of single Vega-Lite specifications alongside metadata to compute responsive multi-column layouts (grid or explicit concatenation) preventing overlaps and styling inconsistencies.
Pydantic
A Python data validation library using standard type hints. Used extensively across all input parsing, LLM structured outputs, and database interaction boundaries to guarantee compile-time and runtime type safety.
Server-Sent Events (SSE)
A unidirectional streaming protocol used over standard HTTP. It enables the FastAPI server to emit live telemetry updates (progress, complete, error) to the React client during multi-stage generation tasks.
Strategy Agent
The entry stage of the generation pipeline. It takes natural language user prompts and database schema context to formulate precise analytical objectives (ChartGoal) outlining the visual approach.
Sub-query Injection
A localized filtering technique used by the fast-filtering engine (POST /dashboard/filter). It wraps base SQL queries inside outer sub-queries appending dynamic WHERE constraints to slice pre-existing dashboard metrics instantly.
Vega-Lite
A high-level declarative visualization grammar. It enables data graphics to be described in JSON format using intuitive visual encoding channels (x, y, color, size, shape) mapping directly to underlying data fields.
Viz Spec Agent
The third stage of the generation pipeline. It accepts raw tabular results alongside targeted objectives to synthesize valid, responsive single-chart Vega-Lite JSON blocks.