Release overview, June 2026
The Curiosity Team
June was about visibility and choice: see exactly what your AI is costing, run several embedding models side by side, let agents collaborate with you mid-task, and see why a result was matched. Here are the June highlights.
Know what your AI costs
The headline this month is full LLM cost and usage observability:
- Every chat completion now records token counts, latency and finish reason.
- A new admin dashboard charts LLM usage over time (tokens, requests and cost).
- Per-model pricing configuration with historical cost tracking, and reference prices that refresh automatically every day from the Curiosity pricing service.
"Finally, an answer to 'what does this cost us?'"
Token counts, latency, finish reasons and per-model pricing come together into one place, so you can attribute spend, spot runaway usage, and plan capacity with real numbers.
Multi-model vector indexing
- Index with several embedding models at once. Per-model vector indexing supports multiple embedding models simultaneously, managed from a new Vector Indexing settings page with chunking options.
- A built-in compact, multilingual embedding model ships as the default, so you get good retrieval without an external provider.
- Import and export embedding vectors as complete bundles.
- Background indexing targets a share of the CPU instead of a fixed thread pool, keeping the workspace responsive while it indexes. Set the target on the new Operate > Indexing CPU page.
Agents that ask
Agents can now pause mid-turn to ask multiple-choice questions, rendered as inline questionnaires. Instead of guessing, an agent can check with you and then continue. That small change makes complex, multi-step tasks far more reliable.
Chats got easier to read, too. Several tool calls in a row now collapse into a single "Tools used" summary, and HTML code blocks render a sandboxed live preview you can flip between "Preview" and "Code".
See why results match
- Similarity results explain themselves. Results from the similarity engine carry a contribution bar that breaks the score down by signal, strongest first, with a consistent color per signal.
- A rebuilt graph explorer for the graph and schema views, with a centered search overlay, node tooltips, and a right-click menu to preview, open, expand and collapse nodes.
Search that understands filters
The search box now understands inline filters like @file, @webpage,
filetype:, ext: and src:, so you can narrow results without leaving the
keyboard. Per-signal weights affect ranking under reciprocal-rank fusion, and
internal node types are excluded from results by default. Search also collects
results several times faster on large indexes and holds onto far less memory.
Operate with less friction
- In-product software updates for Docker deployments, from a new Software Updates page. It can also check an on-premises or mirrored container registry for new builds.
- A Deletion Queue observability page showing nodes scheduled for delayed deletion, and agent-run inspection with reusable admin diagnostic snippets.
- Run scheduled tasks directly from the editor, and a Tasks page that separates custom-code tasks from built-in types.
- Newer OpenAI models (GPT-5.2, plus nano and mini variants) are selected correctly, and custom OpenAI hosts auto-detect their API style.
The little things
- Clear error messages on failed chat replies, with a "Try again" button.
- A built-in local chat-name model generates titles with no external provider.
- Faster loading of large graphs through streaming passes, and lower memory use on big workspaces.
- Provider logos shown in Models & Pricing instead of raw codes.
This overview covers builds v26.6.67170 → v26.7.67930. For the full, build-by-build detail see the changelog.
Read next
Articles on context graphs, enterprise search and industrial AI
