Uncensored AI Production Checklist

Verify Endpoint Compatibility

The first step is confirming that your client library connects to the correct base URL. Our service provides a standard /v1/chat/completions endpoint that mirrors the OpenAI structure exactly. This means you can swap the base URL in your existing SDK configuration without rewriting your core logic.

• Use https://api.hereticproject.com/v1 as your base URL.
• Send requests to POST /v1/chat/completions.
• Query available models via GET /v1/models.

Do not assume other endpoints exist. We do not offer embeddings, images, or audio generation. If your application relies on multimodal inputs, this single-model API may not fit your needs. Stick to text-in, text-out workflows for the best results.

Test Context Window Limits

Our model supports a 100,000-token context window, allowing for extensive conversation history or large document processing. This limit applies to the combined total of input tokens and output tokens. You can retain vast amounts of prior conversation without losing coherence.

When designing your application, ensure you track token usage accurately. If you exceed the limit, the API will return an error. Use the token usage data in the final streaming chunk to monitor your consumption in real-time. This helps you stay within budget and avoid unexpected truncation of your conversation history.

Check Token Output Constraints

Each request can return a maximum of 16,000 tokens of output. If you do not specify the max_tokens parameter, the default limit is 2,048 tokens. This constraint ensures consistent performance and predictable latency for your users.

For long-form creative writing or detailed code generation, always set the max_tokens parameter explicitly. This prevents the model from cutting off mid-sentence if it runs out of space. You can also set a lower limit to control costs if you only need short responses. The API charges based on actual usage, so precise control helps manage your prepaid credit.

Enable Streaming for UX

Streaming via Server-Sent Events (SSE) provides a smoother user experience by delivering tokens as they are generated. This reduces perceived latency, especially for longer responses. The final chunk of the stream includes the total token usage, allowing you to update your billing dashboard instantly.

Implement a streaming client in your frontend or backend service. Handle partial tokens correctly to avoid breaking the UI. This approach is essential for chat interfaces where users expect immediate feedback. Without streaming, users might wait several seconds for a complete response, leading to a poor experience.

Validate Function Calling

The API supports function calling (tools) and tool choice, enabling your application to interact with external systems. You can define tools in the request and let the model decide when to use them. This is crucial for building agents that can fetch data, update databases, or perform calculations.

Ensure your tool definitions match the expected JSON schema. The model will return tool calls in the response, which you must execute and pass back to the conversation. This loop allows for complex, multi-step workflows. Test edge cases where the model might hallucinate tool parameters to ensure robustness.

Set JSON Mode for Structured Data

For applications requiring strict output formats, use the response_format parameter set to {"type": "json_object"}. This forces the model to return valid JSON, which is essential for parsing data programmatically. It reduces the need for complex post-processing or regex extraction.

Validate the JSON output on your side anyway, as no model is perfect. However, this mode significantly increases the reliability of structured data extraction. Use this for generating configuration files, database queries, or API payloads directly from natural language prompts.

Configure Temperature and Top P

Adjusting temperature and top_p controls the randomness of the model. Lower values make the output more deterministic and focused, while higher values encourage creativity and diversity. The uncensored nature of the model means it may produce varied outputs even at low temperatures.

Start with a temperature of 0.7 for a balance of creativity and coherence. Increase it for brainstorming tasks or decrease it for factual retrieval. The top_p parameter works similarly by limiting the probability mass of sampled tokens. Experiment with these settings to find the right tone for your specific use case.

Monitor Usage and Costs

Pricing is transparent: $0.15 per 1M input tokens and $0.50 per 1M output tokens. You pay only for what you use, with no monthly subscriptions. Errors and refusals are free, so you only pay for successful responses. This model is ideal for high-volume applications where cost predictability matters.

Track your token usage through the API responses. Since credits are prepaid and never expire, you can top up as needed. The lack of monthly fees means you avoid paying for idle capacity. Monitor your balance regularly to ensure you have sufficient credits for your peak usage periods.