Skip to content

Releases

v2.2.0

Choose a tag to compare

@github-actions github-actions released this 31 Oct 11:47

Performance & Context Optimization Update

This minor release introduces structural optimizations to improve the pipeline's startup speed and the LLM's extraction accuracy.

🛠️ Changes & Improvements

  • Optimized Label Context: Removed overly specific and redundant target labels. By condensing the label list, I significantly reduced the cognitive load and context window size for the LLM, preventing "overthinking" and improving overall extraction stability.

  • Lazy Loading for Heavy Libraries: Implemented on-demand loading for faster_whisper and gliner2. These heavy dependencies are now initialized only when their specific modules are called, reducing the initial startup time and speeding up the overall pipeline execution.

v2.1.0

Choose a tag to compare

@github-actions github-actions released this 02 Oct 11:24

🚀 Key Features & Performance

  • New Default LLM Engine: Transitioned to Qwen2.5-7B-Instruct-Q4_K_M. Working in tandem with GLiNER and Regex, the pipeline now achieves an outstanding ~91.4% Recall on the ai4privacy benchmark.
  • Hyper-Aggressive PII Detection: The system now successfully extracts complex, multi-word entities (addresses, fragmented dates) and outperforms standard datasets by catching unannotated context-sensitive PII.

🛠️ Architecture & Core Improvements

  • Token Exhaustion Prevention: Redesigned the LLM prompt to implement a "Silent Ignore" rule. The model now strictly outputs JSON only for sensitive words, drastically reducing generation time, eliminating token limit crashes, and preventing JSON parsing errors.
  • Benchmarking Suite: Added an advanced evaluation script to calculate Precision, Recall, and F1-Score locally, including logic to handle partial dataset annotations.

v2.0.0

Choose a tag to compare

@github-actions github-actions released this 31 Aug 17:37

🚀 The Rust Core Engine

This major release marks a structural evolution for WaveRedact. While the AI pipeline remains as accurate as ever, the underlying audio processing engine has been completely rewritten to guarantee enterprise-grade stability, zero-crash memory management, and flawless standalone executables across all platforms.

🦀 The New Rust Audio Engine (waveredact_core)

The previous Python-based audio manipulation (pydub) created multiple memory-heavy copies of the audio files during the redaction process, risking Out-Of-Memory (OOM) crashes on large files.

  • Zero-Copy Memory Safety: Audio manipulation is now handled by a custom-built Rust extension. It uses in-place memory modification directly on the raw bytes, consuming virtually zero additional RAM regardless of the file size.

  • Hardware-Level Generation: The censorship beep (600Hz sine wave) is now generated mathematically at the hardware level per-frame, ensuring perfect alignment without audio glitches.

Seamless Integration: Bridged via PyO3 and Maturin, the Rust core communicates instantly with the Python AI orchestrator.

📜 Enterprise-Ready Audit Trails (Compliance)

To support GDPR and HIPAA compliance workflows, WaveRedact now prioritizes legal safety and corporate accountability.

  • Automated Audit Logs: The tool now automatically generates an Audit Trail file alongside every processed audio in the CLI version, detailing the exact timestamps of redacted entities.

  • Zero-Knowledge Logging: To prevent the log from becoming a data leak itself, WaveRedact logs what was found (e.g., EMAIL, IBAN) and when, but never the underlying raw text.

v1.4.1

Choose a tag to compare

@github-actions github-actions released this 26 Aug 11:41

🐛 CI/CD & Executable Hotfix

This patch resolves a critical build issue affecting the standalone .exe files in recent release (v1.4.0). A conflict in the automated CI/CD environment caused some internal dependencies (questionary, peft) to be excluded from the final packaged binary, resulting in crashes upon launch.
The GitHub Actions pipeline has been completely rewritten using isolated uv virtual environments to guarantee perfectly packaged, ready-to-use standalone binaries for all operating systems.

v1.4.0

Choose a tag to compare

@github-actions github-actions released this 25 Aug 20:57

⚠️ WARNING: The standalone executables in this release are missing dependencies due to a build system bug. Please download version v1.4.1 or higher.

What's New 🚀

🧠 Dynamic LLM Loading (RAM-Aware)

WaveRedact now automatically detects your system's available memory before initializing the Validation LLM, ensuring maximum stability without crashing on lower-end hardware:

  • For systems with <= 8 GB of available RAM: The pipeline now automatically falls back to Gemma-3-4b (Q5_K_XL). This lighter model (~2.64 GB) provides an excellent balance of performance and stability on constrained systems.
  • For systems with > 8 GB of available RAM: The pipeline loads our optimal Qwen2.5-7B-Instruct model for the highest quality PII validation.

⚡ Validation Model Optimizations

To further optimize memory usage and processing speed, the primary validation model has been slightly adjusted:

  • Switched the Qwen2.5-7B quantization from Q5_K_M to Q4_K_M. This reduces the memory footprint to ~4.3 GB while maintaining nearly identical redaction quality and reasoning capabilities.

v1.3.0

Choose a tag to compare

@github-actions github-actions released this 24 Aug 11:19
6942376

🚀 Precision Tuning & Prompt Optimization

In this update, I focused on refining the synergy between the GLiNER extraction engine and the LLM validation layer to prevent data leakage and improve overall processing speed.

✨ Key Changes & Improvements

  • Smart LLM Delegation (Threshold Tuning): During extensive benchmarking, I noticed that local LLMs sometimes tend to "over-correct" and drop valid PII redactions, risking data leaks. To solve this, I introduced a low confidence threshold at 0.75.
    • If GLiNER is enough confident (> 0.75) that a word is sensitive, it bypasses the LLM entirely.
    • The LLM is now exclusively triggered as a fallback to resolve truly ambiguous cases (<= 0.75). This change improved recall by 8–10%.
  • Prompt Overhaul for Small Models: The internal prompts have been rewritten. I removed repetitive instructions and restructured the logic (Role ➔ Rules ➔ Output Format ➔ Data) to be perfectly optimized for small-parameter models (1B to 7B). This prevents context-loss, speeds up inference, and drastically improves the model's adherence to the strict JSON output schema.

Update your local environment to experience a faster and more secure redaction pipeline! 🌊

v1.2.0

Choose a tag to compare

@github-actions github-actions released this 20 Aug 15:42

Release v1.2.0 🚀

This release introduces critical bug fixes that drastically improve the reliability of the Data Privacy Pipeline, specifically focusing on the local LLM generation stability and the accuracy of the Regex Extraction phase.

🐛 Bug Fixes

  • Fixed LLM Hallucination Loop (KV Cache Quantization Removed): Resolved a critical issue where the local Llama server would generate endless sequences of numbers or gibberish (hallucination loop). This was caused by the q4_0 4-bit KV Cache quantization conflicting with modern LLM architectures (like Qwen 2.5), which are highly sensitive to precision loss in their attention mechanisms. The server now defaults to f16 (16-bit float) for the KV Cache, ensuring perfect generation stability and JSON structured outputs without breaking the context window.
  • Dynamic Target Enforcement in Regex Extractor: Fixed an issue where the RegexExtractor was unconditionally building and applying regular expressions for all default data types (Email, IBAN, Phone, etc.) regardless of the user's active Level configuration. The extractor now strictly filters its regex patterns, dynamically mapping and executing only the specific labels requested by the user, completely eliminating Regex-based False Positives for unselected entities.

v1.1.0

Choose a tag to compare

@github-actions github-actions released this 18 Aug 10:33
530063a

🚀 Custom PII Labels & Targeted Redaction

This release introduces a granular level of control over audio data censorship. Alongside the classic presets (base, medium, total), you can now tell WaveRedact exactly what to look for and redact using custom JSON profiles!

✨ New Features & Improvements

  • Tailored Targeting (Custom Labels): Do you need to redact only emails and passwords while leaving names and addresses untouched? Now you can. Just create a simple JSON file containing an array of labels (e.g., ["email", "password", "iban"]) and feed it to the CLI.
  • Strict Validation System: I've implemented a validation system for custom inputs. WaveRedact will check in real-time that your JSON is properly formatted and, most importantly, that the requested labels belong to the domain supported by the specific GLiNER PII model.

v1.0.0

Choose a tag to compare

@github-actions github-actions released this 08 Aug 18:47
b9378a6

What's Changed

  • Added tests and covered unavailable port problem in llama_server

v0.10.2

Choose a tag to compare

@github-actions github-actions released this 08 Aug 11:25
8ea3a49

Minor update

  • Fixed ux and project folders