Pnyx launches OpenAI‑compatible Pnyx‑Privacy‑Filter v1 to redact PII locally before prompts

Pnyx AI said on May 22 it had released Pnyx-Privacy-Filter v1, a software stack that sits in front of a language model and removes personally identifiable information before prompts are sent onward. The company described the system in posts on X as an OpenAI-compatible proxy for enterprise deployments. Pnyx’s public GitHub repository says the stack detects and redacts PII before messages reach the model, then de-anonymizes responses before returning them to the client. The code published by Pnyx shows the product is built around two main pieces: a Triton Inference Server deployment for the underlying privacy model and a “PrivateChatManager,” or PCM, that handles the redact-forward-de-anonymize flow. The repository says PCM exposes an OpenAI-compatible API and, in its current form, supports `v1/chat/completions`, with broader API coverage planned later. ### What exactly is Pnyx shipping here? Pnyx’s GitHub repository describes Pnyx-Privacy-Filter as a “privacy-filtering stack for LLM deployments.” The README says the system uses a containerized version of OpenAI’s Privacy Filter model, served through Triton, while Pnyx’s own proxy layer manages request handling and response reconstruction. (github.com) OpenAI said last month that its Privacy Filter is a bidirectional token-classification model for personally identifiable information detection and masking in text. (github.com) OpenAI’s repository says the model is intended for high-throughput sanitization workflows and can run on premises, while Pnyx’s repository says its stack includes derivative works of `openai/privacy-filter` under the Apache 2.0 license. ### How does the prompt redaction flow work? (github.com) Pnyx’s README says PII is detected and redacted before a message reaches the language model, and that responses are de-anonymized before being returned to the client. The repository says PCM maintains a placeholder map for the full session, so if a later response refers back to an already redacted entity, the model sees the placeholder rather than the raw text. The same README says that session tracking is designed to improve consistency across turns and tool calls. (openai.com) Pnyx also says the proxy includes basic support for tool or function responses, though it notes that structured-data handling can be inconsistent because the underlying model was not specifically trained for that use case. ### How close is it to the OpenAI API? The repository published by Pnyx says PCM is “OpenAI-compatible,” but it also specifies that compatibility is currently limited to `v1/chat/completions`. (github.com) The README says that support is expected to be extended in future versions. The examples folder in the repository includes a Docker Compose setup, and Pnyx instructs users to configure environment variables and point the proxy at an upstream language model endpoint. (github.com) The sample stack includes a health check against a local PCM service running on port 8080. ### What are the stated limits? Pnyx’s README says the current system is text-only and does not redact PII embedded in images, PDFs or other file formats. The company also says tool and function response support is still basic, and points users to OpenAI’s own documentation for bias risks and other model limitations. (github.com) OpenAI’s repository says its Privacy Filter model has 1.5 billion parameters in total, with 50 million active parameters, and supports a 128,000-token context window. (github.com) Those are the model characteristics Pnyx is building around, according to the repository and license notice. ### Where can users inspect the release? Pnyx’s public repository is live on GitHub under `pnyxai/pnyx-privacy-filter`, with Docker files, a model repository, example configurations and the PrivateChatManager source code. (github.com) The repository page showed one public commit and no packaged GitHub release entry when checked on May 22. OpenAI’s upstream `openai/privacy-filter` repository is also public, and Pnyx’s README directs users to that project for underlying model limitations and attribution details. (github.com) The next concrete step for users is in Pnyx’s examples directory, where the company provides the `.env` setup and Docker Compose commands to run the stack locally. (github.com)