Playground TextSynth

Provides a single REST API endpoint that abstracts over multiple language models (GPT-3, GPT-J, Mistral) with consistent request/response schemas, eliminating the need to manage separate API keys or learn different SDKs per provider. Requests specify the target model as a parameter, and responses include token counts and model metadata, enabling programmatic model selection and cost tracking without vendor lock-in.

Implements granular, pay-as-you-go billing where each API request returns exact token counts (input and output tokens separately) and charges are calculated at request time without subscription minimums or monthly commitments. The pricing is published per-model and per-token-type, allowing developers to predict costs before making requests and optimize for cost-per-task rather than fixed monthly fees.

Provides a web-based interface where developers can enter a single prompt and execute it against multiple models (GPT-3, GPT-J, Mistral) simultaneously or sequentially, displaying outputs in parallel columns with metadata (tokens used, latency, model name) for direct visual comparison. The UI supports adjustable hyperparameters (temperature, top_p, max_tokens) that apply across all selected models, enabling controlled A/B testing of model behavior on identical inputs.

Supports HTTP streaming (Server-Sent Events or chunked transfer encoding) for text completion requests, returning tokens incrementally as they are generated rather than waiting for the full response. This enables real-time display of model outputs in client applications, reducing perceived latency and allowing users to see partial results while generation is in progress, with each chunk including token metadata for cost tracking.

Accepts temperature, top_p, top_k, and max_tokens parameters in API requests with model-specific valid ranges enforced server-side. The API validates parameters against each model's constraints (e.g., GPT-3 supports temperature 0-2, GPT-J supports 0-1) and returns errors for out-of-range values, preventing silent failures or unexpected behavior from invalid configurations.

Designed as a stateless REST API where all functionality (model selection, parameter tuning, streaming) is available via HTTP endpoints, with the web playground UI as an optional thin client that consumes the same API. This architecture enables developers to build custom interfaces, integrate into existing workflows, or use the API directly without relying on the web UI, and allows the API to evolve independently of UI changes.