PhariaAI v1.260900.0

Opt in to the OpenAI-compatible GET /models List Models format on the legacy (/v1 and no-prefix) route. When enabled, the admin config-map view of /models is removed; administration stays on /model-packagesand /checkpoints. openaiCompatibleModelsEndpoint: bool

Added support for native tool calling.

Fixed tool call history being dropped from A2A conversation context, which caused the model to lose tool context after several turns and stop invoking MCP tools.

Fixed incognito stages not being cleaned up after a force-quit.

Improved MCP logging and log-level hygiene. Potentially sensitive data (tool parameters, error content, LLM buffer data) was moved from INFO to DEBUG level.

v0.9.0 removes the Parsy chart-polish pass. Chart output now comes from the structured inference path.

Hybrid search. Search stores can combine BM25 lexical and dense vectors. Fusion strategies are convex, "anchored", and a lexical floor for sentence-like queries. New stores of this kind default to convex fusion. The fusion param prefetch_limit controls prefetch depth. The MCP search tool supports reranking and metadata filters.

Encryption and downloads. Data is encrypted at rest in Postgres and S3. File content is served through signed download URLs, including through the CDN. Compliance docs cover access control, retention, and audit logging.

Export and clone. Export a search store as a downloadable zip, including files.jsonl and vectors.parquet. Tenant-wide export covers GDPR Art. 15 and 20. Clone a search store. Exports and clones are hidden from workflow and file listings unless the caller sets include_managed.

New file formats. Word, Excel, PowerPoint, OpenDocument, and EPUB files are ingested. Audio files are transcribed, and transcripts appear as pages and searchable chunks. PDF text extraction and rendering improved, with per-page routing between text extraction and OCR.

File metadata. Upload accepts user metadata, which lands on chunks for filtering. A metadata update endpoint cascades to chunks. String metadata indexes support prefix matching. Invalid filter values return 400 instead of 500.

Run behavior. One failed file no longer fails the whole batch. Failed files are listed per file in the run result. Cancelled runs report as cancelled, not failed, and the SDK stream ends on cancel.

Integrate Narra Voice API

New endpoint for feedback export

Conversation metrics: added tracking of unique user interactions

Conversation export: new streaming endpoint for exporting anonymize conversations

Final answer customization: added support for custom instructions in the final answer prompt

Citation fixes

Improved error responses & logging

Added Stille 3.2 translation model with simple-split and server-sent event streaming

Added new prompt mechanism for Stille 3.1 and 3.2

Added asynchronous upload configuration support

Added monitoring and logging improvements

Monitoring and logging improvements

Fixed error code mapping and timeout handling

Added support for asynchronous upload configuration

Fixed error code mapping and request timeout handling

Fixed pagination order and filtering in feedback endpoint

CVE fixes across all services