Magic Potion

Provides a drag-and-drop node graph interface for constructing AI prompts without writing code. Users connect visual nodes representing prompt components (input variables, instructions, conditionals, output formatting) into a directed acyclic graph that compiles to executable prompt chains. The editor likely uses a canvas-based rendering system (WebGL or SVG) with node serialization to JSON/YAML for persistence and execution.

Abstracts away provider-specific API differences (OpenAI, Anthropic, Cohere, local models) behind a unified execution layer. The editor compiles visual prompt graphs into provider-agnostic intermediate representation, then routes execution to the selected provider's API with automatic parameter mapping (temperature, max_tokens, stop sequences). Likely implements adapter pattern with provider-specific SDKs or REST wrappers.

Provides centralized repository for storing, organizing, and reusing prompt templates across projects. Implements tagging, search, and categorization for discovering templates. Supports template inheritance where specialized prompts extend base templates, reducing duplication. Includes template metadata (description, author, tags, usage examples) and version control. May support community sharing or private team libraries.

Maintains version history of prompt graphs with branching support, allowing users to create variants and run A/B tests comparing outputs. The system likely stores graph snapshots with metadata (timestamp, author, description), implements diff visualization for prompt changes, and provides statistical comparison tools (win rate, average quality scores) across test variants. May integrate with evaluation frameworks to automate quality assessment.

Supports parameterized prompts where variables (e.g., {{user_input}}, {{context}}) are substituted at execution time from multiple sources: form inputs, API responses, database queries, or file uploads. The system implements a template engine (likely Jinja2-style or custom) that handles type coercion, escaping, and conditional inclusion of variables. Context injection allows pulling external data (documents, knowledge bases, API results) into prompts before execution.

Records every prompt execution with full context: input variables, selected model, parameters, output, latency, token usage, and cost. Stores execution logs in a queryable database with filtering by date, model, prompt version, or outcome. Provides audit trail for compliance and debugging, with optional integration to external logging services (DataDog, Splunk). May include execution replay functionality to reproduce specific runs.

Enables multiple users to edit the same prompt graph simultaneously with real-time updates, conflict resolution, and change notifications. Likely implements operational transformation (OT) or CRDT (Conflict-free Replicated Data Type) for concurrent editing, with WebSocket-based synchronization. Includes user presence indicators, comment threads on nodes, and role-based access control (view, edit, admin).

Provides framework for defining and running custom evaluation functions against prompt outputs. Users can write evaluation logic (e.g., 'check if output contains required keywords', 'score relevance 1-5') as code or visual rules, then batch-run evaluations across test datasets. Integrates with common evaluation libraries (RAGAS, DeepEval) or allows custom metric definitions. Results displayed as pass/fail rates, score distributions, and failure case analysis.

Packages prompt graphs as deployable artifacts (Docker containers, serverless functions, REST APIs) that can be deployed to production infrastructure. Handles API endpoint generation with request/response schema validation, rate limiting, authentication, and monitoring. Likely supports multiple deployment targets: cloud platforms (AWS Lambda, Google Cloud Functions), container registries, or self-hosted servers. Includes rollback and canary deployment capabilities.

Connects prompt graphs to external knowledge sources (documents, databases, vector stores) for context retrieval. Implements RAG pattern where user queries are embedded, matched against knowledge base, and relevant documents injected into prompts before LLM execution. Supports multiple knowledge source types: uploaded documents, web search, API endpoints, or vector databases (Pinecone, Weaviate). Includes chunking, embedding, and retrieval configuration.

Supports conditional execution paths in prompt graphs where different prompt branches execute based on input conditions or intermediate results. Implements if/else nodes, switch statements, and loop constructs for iterative prompt execution. Conditions can be based on input variables, LLM outputs, or external API responses. Enables complex workflows like 'if input contains question, use Q&A prompt; else use summarization prompt'.