Intelligent Research & Polished Experience

AI Chat

• Evidence-First Citations — Chat responses now cite specific evidence from your sources with claim-level binding, integrity scoring, and reason codes.
• Source Inspector — Split-panel inspector lets you view citation evidence and source content side-by-side during conversations.
• Umbrella Thinking Steps — Visible AI reasoning stages with weighted smoothing show how the AI arrives at its answer.
• GPT-5.4 Models — Access to latest GPT-5.4 models via a redesigned ghost-button model selector.
• Two-Tier Model Selector — Quickly switch models with a popover showing your recent selections.
• Redesigned Submit Button — New outline-to-fill morph animation with send icon.

Workflows

• Agentic Auto-Execution — Multi-step AI workflows now run autonomously with a chat UI refresh, pausing for approval when needed.

Sources

• Mistral OCR for PDFs — PDF extraction now uses Mistral OCR for significantly better text quality from scanned documents.
• Content Sanitisation — New sanitisation step in the ingestion pipeline improves source quality before embedding.
• Map-Reduce-Map Extraction — Entity extraction now uses a scalable map-reduce-map pipeline for more reliable knowledge graph building.

Knowledge Graph

• Brain Map Layout — New organic layout option for the knowledge graph canvas, offering a more natural spatial arrangement of concepts.

Interface

• OnsombleAvatar — New animated AI avatar with mode-tuned color system and choreographed state transitions that respond to AI activity.
• Nano-Banana-2 — Upgraded image generation model for AI avatar creation.

AI Chat

• Faster Input Response — Isolated input state prevents unnecessary ChatPanel re-renders for smoother typing.
• Unified Onsomble Voice — Final generation now uses a consistent Onsomble voice for more natural, cohesive responses.

Search & Retrieval

• Batch Embeddings — All queries are now batch-embedded with multi-query vector search for faster retrieval.
• Adaptive Comparison Retrieval — Retrieval dynamically adjusts based on query complexity.
• Speculative Retrieval — Classification and retrieval now overlap, reducing time-to-first-result.
• Confidence-Based Rerank Skip — High-confidence results bypass reranking for faster answers.
• Optimised Query Classification — Deterministic routing, parallelised classify + expand, and per-node model constants reduce classification overhead.
• Evidence Compression — Comparison executor compresses evidence for more efficient context usage.
• Observability — Granular Langfuse spans across graph search and ingestion pipelines for better performance monitoring.

Interface

• Refined Avatar Colors — Light mode avatar colors tuned with per-mode curve control for a more polished appearance.

AI Chat

• Fixed agentic resume progress streaming not displaying correctly
• Fixed conversation titles being truncated mid-word
• Fixed generated titles not persisting on agentic HITL interrupt
• Fixed score-type-aware full-document fallback for RAG retrieval

Sources

• Fixed download failures now correctly classified as retryable
• Fixed Neo4j timeout on large knowledge graph builds with added ingestion resilience

Observability

• Fixed dev queue namespace isolation for tracing
• Fixed rerank diagnostics staying within parent trace