Skip to main content

PhariaAI v1.250600.0 – Release Notes

Release Versionv1.250600.0
Release DateJune 4, 2025
AvailabilityOn-premise and hosted

PhariaAssistant​

In this release, we introduce key improvements that make interaction more fluid and workstreams more organised. Together, these updates focus on one major goal: boosting user efficiency through real-time responsiveness and structured dialogue management.

Real-Time Token Streaming: Faster Answers When Every Second Counts​

Token Streaming enables real-time generation of responses in PhariaAssistant, allowing you to start reading even while the AI formulates its output — you no longer need to wait for full replies to load. This upgrade delivers a smoother, more responsive experience, which is particularly useful in time-sensitive workflows.

Key Benefits​

  • Immediate access to insights: Stream tokens in real-time and begin reading output instantly.
  • Improved user experience: Reduce wait time and improve decision-making speed.
  • Frictionless upgrade: No configuration needed — streaming is now built into PhariaAssistant.
note

This feature is currently only available in the Chat interface; other endpoints are coming soon

Local Chat History: Structure Your Work, Stay in Context​

With Local Chat History, PhariaAssistant now supports managing multiple, topic-specific conversations in a single workspace. Every session is stored directly in the browser, so users can pick up exactly where they left off — whether switching projects or revisiting past queries.

Key Benefits​

  • Stay organised: Manage tasks and topics in separate conversations.

  • Resume work faster: Access previous chats without re-asking questions or re-uploading files.

  • Keep relevant context at hand: Maintain continuity and avoid redundant interactions.

note

Local Chat History is stored per browser and device. It does not sync across devices or include cloud backups.

Upgrade Considerations: Helm Chart Configuration Changes​

If you are upgrading from a previous version of PhariaAssistant, existing Helm chart configurations must be updated. Specifically, environment variables previously defined under pharia-assistant-api must now be reassigned under pharia-chat.

Variable Mappings​

Old KeyNew Key
pharia-assistant-api.env.DOCUMENT_INDEX_FILE_UPLOAD_NAMESPACEpharia-chat.env.values.PHARIA_CHAT_DOCUMENT_INDEX_FILE_UPLOAD_NAMESPACE
pharia-assistant-api.env.RETRIEVER_QA_INDEX_NAMEpharia-chat.env.values.PHARIA_CHAT_RETRIEVER_QA_INDEX_NAME
pharia-assistant-api.env.QA_MODEL_NAMEpharia-chat.env.values.PHARIA_CHAT_DEFAULT_MODEL

Example Migration​

Before (previous Helm chart):

pharia-assistant-api:
  env:
    QA_MODEL_NAME: "llama-3.3–70b–instruct"

Now (post-upgrade configuration):

pharia-chat:
  env:
    values:
      PHARIA_CHAT_DEFAULT_MODEL: "llama-3.3–70b–instruct"

Ensure that you apply the same transformation for all three variables if they are present in your current deployment.

note

Important: These changes are critical for customers with an existing installation who want to upgrade to v1.250600.0. Failure to update Helm values will result in misconfiguration or failed runtime behaviour.

PhariaCatch​

[New Release] Introducing PhariaCatch Initiate​

Capture high-quality domain expertise by:

  • Adding multiple use cases
  • Managing dataset creation workflow
  • Assessing data quality
  • Exporting datasets to be used by AI engineers to build optimised AI systems with precision

With PhariaCatch Initiate, domain experts can now go beyond single-use annotation. This new feature:

  • Enables the creation of multiple use cases within the same platform
  • Supports collaborative dataset building
  • Delivers insights on data quality, empowering AI teams to work smarter and faster

This update simplifies project setup, improves dataset integrity, and reduces manual coordination.

Key Benefits​

  • Build datasets without annotator bias: Use cross-annotation and consistency checks to quality-check the data by multiple data experts.
  • Support multiple use cases: Set up and manage projects across different AI needs within one interface.
  • Streamline team collaboration: With clear role assignment (such as project manager or annotator) and access control, ensure clear management of dataset creation, quality management, and export permissions.
note

PhariaCatch currently supports classification use cases only.

Upgrade Considerations - Helm Chart Configuration Changes​

If you are upgrading from a previous PhariaAI version that did not include PhariaCatch, a new PostgreSQL database must be configured for PhariaCatch. This database is required for PhariaCatch to store annotation projects, datasets, and user collaboration data.

PostgreSQL Configuration​

PhariaCatch requires its own dedicated PostgreSQL database instance. You must configure the following in your Helm chart values:

pharia-catch-api:
  postgresql:
    enabled: true
    auth:
      password: "your-catch-db-password"
      # Alternatively, use an existing secret
      # existingSecret: "pharia-catch-postgresql-secret"

Example Configuration​

For new installations with built-in PostgreSQL:

pharia-catch-api:
  enabled: true
  postgresql:
    enabled: true
    auth:
      password: "your-catch-db-password"

For installations using external PostgreSQL:

pharia-catch-api:
  enabled: true
  postgresql:
    enabled: false
    auth:
      existingSecret: "pharia-catch-postgresql-secret"

PhariaStudio​

In this release, we introduce two powerful features designed to enhance team collaboration and improve observability of AI systems in production environments.

Project-Level Collaboration in PhariaStudio​

Maintain multiple AI projects within a workspace, and collaborate with team members directly within projects while maintaining clear access boundaries and role-based permissions.

With this release, you can add team members to individual projects, assign them roles (such as project owner or member), and ensure strict data separation between different AI initiatives - all from a single workspace.

Key Benefits​

  • User access control: Assign project-specific roles to limit data access and reduce security risks.
  • Data separation: Prevent cross-project data exposure by isolating datasets and members per project.
  • Deployment isolation: Each project can manage its own deployment lifecycle, ensuring that updates in one project do not affect others.
note

You are currently limited to a single workspace.

Real-Time Trace Collection for PhariaKernel Skills​

Capture real-time traces from PhariaKernel Skills used in production that follow the OpenTelemetry (OTel) standard for trace management.

Previously, you could collect only metrics and logs. But in this release, you can now collect and analyse detailed trace information from Skill execution in runtime environments. This empowers you to obtain greater AI system observability and improve debugging.

Key Benefits​

  • Production-grade trace visibility: Monitor real-time behaviour of PhariaKernel Skills and services.
  • OTel-compliant trace management: Use standard tooling and export traces to preferred observability platforms.
  • Improved debugging: Gain insight into how Skills behave during live execution to refine and troubleshoot complex systems.
note

Real-time export of traces to external systems is not yet supported.

PhariaOS​

[New Release] Introduction: Foundational Capabilities for Model Control, Efficiency, and Reach​

In this release, we introduce a set of foundational capabilities that guide you from greater control over your AI infrastructure to broader reach across your organisation.

This release reflects four product pillars:

  1. Control: Dynamic Model Management (you control what runs).
  2. Efficiency: Shared Inference (you optimise how it runs).
  3. Intelligence: Steering Concepts (you guide how it thinks).
  4. Reach: Transcription (you expand where it listens).

Dynamic Model Management: Deploy and operate model without touching code​

Dynamic Model Management enables you to deploy and manage Aleph Alpha-supported and open-source language models - without touching infrastructure. The feature provides a unified interface to explore, install, and control LLMs with ease, all within PhariaOS.

Key Benefits​

  • One-click deployment: Manage commercial and open-source models through a streamlined UI.
  • Fewer deployment errors: Eliminate manual Helm edits and config mismatches.
  • Faster time to production: Skip complex setup steps and accelerate model rollout.
note

Adapter-based models and finetuning not yet supported.

Shared Inference: Reduce costs and minimise overhead by sharing GPU resources across isolated deployments​

Shared Inference is a new capability that helps reduce GPU costs and complexity by enabling multiple PhariaAI instances to securely share a single inference stack, with authenticated access and full isolation for each instance.

Key Benefits​

  • Lower GPU overhead: Share GPU capacity across deployments without duplicating inference infrastructure.
  • Secure by design: Each instance remains isolated with its own IAM, ensuring compliance and control.
  • Operational simplicity: Deploy once and scale across departments or business units with ease.
note

This feature is only available with the PhariaInference API; UI support is coming soon.

User-Defined Steering Concepts: Save time, compute, and prompt tokens by guiding model behaviour directly at inference​

User-Defined Steering Concepts is a new capability that allows you to customise model behaviour by defining your own steering concepts through the API or Python client.

Key Benefits:​

  • Faster iteration: Define and refine model behaviour without retraining or complex prompt design.
  • Granular control: Tailor steering concepts to your use case and deploy them programmatically.
  • Token efficiency: Save prompt space by steering responses without inflating the context.
note

This feature is available with the API and Python client only, with support for Llama-3.1-8B-Instruct.

Transcription via PhariaInference API: Expand how the model understands with the updated Transcription Worker​

The transcription functionality has been upgraded and made accessible through the PhariaInference API, replacing the internal-only pharia-transcribe.

This update enables external applications to initiate transcription tasks with support for larger files, asynchronous job queuing, enhanced timestamping, and authentication-based reporting.

Key Benefits​

  • Streamlined Integration: A public, authenticated API endpoint enables easy embedding of transcription workflows into enterprise systems and external applications.
  • Enterprise-Ready Scale & Reliability: Support for files up to 200 MB and a queuing backend ensures stable processing even under heavy loads.
  • Enhanced Transcript Usability: Sentence-level timestamps improve accuracy, readability, and facilitate downstream applications like summarisation and quotation.
note

Transcription tasks are currently processed sequentially per worker, which may impact throughput at scale. Real-time or streaming transcription is not yet supported.

Other updates​

  • Bug fixes & improvements
  • Security enhancements
  • Minor performance and stability fixes