Changelog for AI Assistant

• Added
  Adds Studio, a built-in surface served by the AI Assistant server for configuring an assistant (model, system prompt, and skills) and chatting with a document to test it. (#55385)
• Added
  Adds a top-level agents management page (served at /agents) for browsing, editing, versioning, and deleting saved agents, including a settable agent identifier, and turns the Studio sidebar into a quick-update surface for the selected agent. (#56369)
• Added
  Adds import and export of agent configurations on Studio’s agents page: Export the selected agent as a portable JSON file, and import a file as a new agent or a new version of an existing one. Exported files never contain provider credentials, and saving an agent that embeds a secret returns a 400: provider credentials (apiKey, Bedrock access keys), HTTP MCP request headers, or an MCP url that carries a username, password, or query string. Send secrets with each run request instead, or set provider credentials in the server environment. (#56406, #57489)
• Added
  Adds a config builder to the Studio agents page, letting you describe how an agent should behave differently and have the matching system prompt and skill changes drafted into the form, where you can review them before saving. (#56759)
• Added
  Adds saved agents, letting you persist a custom agent configuration on the server and invoke it by ID through the v2 API, with immutable version snapshots so an earlier configuration can be inspected or restored. (#55431)
• Added
  Adds subagents: named, separately configured agents that the main agent can delegate a task to, each with its own system prompt, tools, and tool-approval policy, and running in isolation so the main agent sees only the final answer. Subagents are configured with the subagents field on an AIAv2 run request. (#55651)
• Added
  Adds an optional toolCalling field to a skill, declaring whether its content is written for direct or code mode. A tagged skill is served only to requests running in that mode, so two variants can share one name; an untagged skill is served in every mode, leaving existing configurations unchanged. (L#AI-24)
• Added
  Adds the allowed_documents JWT claim for per-document layer access control, mirroring Document Engine: either any (unrestricted) or a set of { document_id, layer } grants, where an omitted layer means the default layer and any unlisted layer is rejected. Layer scoping is enforced across all access paths, including wildcard agent runs and document-tool/MCP access. (J#AI-164)
• Added
  Adds an authenticated POST /api/license/refresh dashboard endpoint that reactivates the license from the licensor on demand, so a stale-but-valid cached license key can be refreshed on a running instance without waiting for the periodic licensor ping or restarting. (J#AI-333)
• Fixed
  Fixes an issue where a request with a valid responseSchema returned no structuredResponse on the default model, responding with HTTP 200 and a prose answer only. (J#AI-488)
• Fixed
  Fixes an HTTP 500 from the v1 GET /client/api/v1/sessions/:sessionId endpoint for a session whose history was produced through the v2 inference API. (#55630)
• Fixed
  Fixes duplicate application rows created when server replicas start against an empty database at the same time, which could make license reads nondeterministic and trigger needless licensor reactivation. (#55686)
• Fixed
  Fixes an issue where an in-progress AI Assistant turn could be lost when the client disconnected. The prompt is now persisted on receipt and the streamed response is flushed while generating and immediately on disconnect, so reopening the session restores the latest message instead of losing it. (J#AI-112)

• Fixed
  Fixes AI Assistant failing on current Anthropic models by no longer forcing an unsupported extended-thinking configuration. Model parameters are now a straight configuration pass-through and are no longer inferred from the model name; a model configured without an explicit temperature now uses its provider’s default instead of a forced value of 0. (J#AI-495)
• Fixed
  Fixes an HTTP 500 from the Document Compare change analysis and tagging endpoints when an Anthropic model with extended thinking is used. (J#AI-494)

• Changed
  Changes the default chat LLM to openai:gpt-5.4-mini, which is also used as the fast-fallback model. (#54756)
• Fixed
  Fixes an issue where an expired license key cached in the database could not be reactivated. When the cached key fails validation, the renewed license is now fetched from the licensor immediately, instead of only on a later scheduled ping. (J#AI-334)
• Fixed
  Fixes an issue where streamed tool approval requests could execute ask-gated tools instead of pausing for approval. (J#AI-483)
• Fixed
  Fixes an issue where the X-Pagination-Total header for thread search could be smaller than the number of results returned in the page when threads were created concurrently. (#54750)
• Fixed
  Fixes the redaction analyze REST API documentation to document the stagedActions response field, matching the endpoint’s actual output. (J#AI-476)

• Removed
  Deprecates the modelOverrides request field and agent_configuration.model_overrides JWT claim allowlist. Legacy requests continue to work via auto-translation into the new modelServices.models shape. (J#AI-465)
• Added
  Adds a configurable reranker-llm model label, allowing a dedicated model for LLM-based search result reranking instead of using the default LLM. (#51885)
• Added
  Adds per-label and per-provider allowlists to the agent_configuration.model_services JWT claim, letting operators gate which models each label may use and which providers a caller may supply per request, rather than the previous all-or-nothing Boolean. (J#AI-465)
• Added
  Adds the modelServices request field and agent_configuration.model_services JWT claim for per-request LLM provider and model overrides, applied per-call only and never persisted. (J#AI-465)
• Changed
  Improves chat agent token efficiency, reducing token usage by 40 percent per conversation turn. (#51907)
• Changed
  Improves semantic search recall by over-retrieving candidates from pgvector before LLM reranking, improving correctness on multidocument benchmarks. (#51607)
• Changed
  Improves token efficiency by compressing system prompts and context injection templates. (#51468)
• Fixed
  Fixes an issue where missing MCP servers emitted warnings instead of errors, ensuring failures are surfaced through the API. (J#AI-419)
• Fixed
  Fixes an issue where AI Assistant document tools were one page off, causing annotations and page reads to land one page away from the page the user referred to. (J#AI-463)

• Added
  Adds v2 model configuration with label-based model selection, enabling optimization and subgraph model usage in the future. (#50367)
• Added
  Adds session_ids JWT claim enforcement across all AIAv2 REST API endpoints. (#50699)
• Changed
  Updates the @langchain/anthropic dependency to improve token limit handling. (#50692)
• Fixed
  Fixes an issue where tasks running for more than 60 seconds could end up in an inconsistent state. (#50895)

• Fixed
  Fixes an issue where the PGSSL environment variable was ignored, causing connection failures to SSL-required PostgreSQL databases. (#50374)

• Fixed
  Fixes CVE-2025-66414 and CVE-2026-0621 vulnerabilities. (#50315)
• Fixed
  Fixes an OpenAI model issue when multiple tool calls are made in parallel. (#50316)

• Changed
  Introduces AI Assistant v2, supporting agents with tool use and reasoning traces.
• Changed
  Improves system prompt guards to prevent malicious attacks and prompt injections. (#49600)
• Changed
  Updates langgraph, langchain, and zod dependencies. (#49509)
• Fixed
  Fixes a batch size issue when embedding large documents using OpenAI models. (#50159)
• Fixed
  Fixes a bug in which the AI assistant was unable to provide summaries. (J#AI-283)
• Fixed
  Fixes duplicate human messages being appended to requests. (#49622)
• Fixed
  Fixes feature enablement in evaluation mode for all features. (#49408)

• Added
  Adds support for bearer token authentication in addition to the existing Token token= format. (#48812)
• Added
  Adds support for password-protected PDFs to AI Assistant. (J#AI-28)
• Changed
  Improves database configuration by only requiring the DATABASE_URL environment variable. (#48886)
• Changed
  Changes handling of the empty document_ids array in JWT claims so that it restricts access to no documents instead of all documents. (#49082)
• Changed
  Improves ordinal page references in redaction queries. (#48367)
• Fixed
  Fixes an issue ingesting documents when they contain no changing or permanent IDs. (#48503)
• Fixed
  Fixes an issue where models without temperature support fail to initialize. (#48496)

• Added
  Adds support for generating answers using information from multiple documents. (J#AI-173)
• Changed
  Improves Document Comparison category hits, avoiding some uncategorized changes. (#47978)
• Changed
  Improves document retrieval through chunking documents by paragraphs, tables, and key-value pairs. (J#AI-184)
• Fixed
  Fixes an issue where rate limiting may not work correctly in some scenarios. (J#AI-209)
• Fixed
  Fixes an issue where the rate limiting error message may incorrectly refer to ‘messages per day’, even when you configure a different time period. (J#AI-200)

• Changed
  Improves case insensitivity when identifying redaction terms. (#47440)
• Changed
  Improves ingestion performance for all documents. (#47650)
• Changed
  Improves the size of the Docker image by optimizing its internal structure. (J#AI-181)
• Changed
  Improves document retrieval across single-document and multidocument queries. (#47583)
• Fixed
  Fixes bundle ID checking for origin checks from iOS. (#47784)
• Fixed
  Fixes an error message about an invalid document index if multiple documents are ingested simultaneously. (J#AI-191)
• Fixed
  Fixes an issue that may happen when ingesting a document with multiple layers from Document Engine. (J#AIA-178)

• Added
  Adds AI-powered text comparison. (J#AI-171)
• Changed
  Improves term detection in redaction when content is short. (#47404)
• Changed
  Improves term detection in the redaction agent. (J#AI-149)
• Changed
  Optimizes the loading of documents from Document Engine by avoiding unnecessary ingestions. (J#AI-163)

• Changed
  AI Assistant no longer requires Document Engine to function and can now be used by other Nutrient SDKs, including the Web SDK. (J#AI-75)
• Added
  Adds a new CONFIG_DIR environment variable that states where the service-config.yml is located in the container. (J#AI-119)
• Added
  Adds new Server API to ingest documents from Document Engine. (J#AI-110)
• Changed
  Improves REST API error messages by including detailed information about failing paths. (#46262)
• Changed
  Improves the redaction workflow by interpreting natural language page range instructions and prompting follow-up actions for clarification when page ranges are unspecified. (J#AI-81)
• Fixed
  Fixes CVE-2024-21538 by removing external dependency on express-requests-logger. (J#AI-108)
• Fixed
  Fixes an issue where the document IDs of some documents may be incorrectly truncated. (J#AI-123)
• Fixed
  Fixes an issue where uploading documents with lots of pages may fail unexpectedly. (J#AI-128)

• Fixed
  Fixes OpenAI-compatible model-provider usage when no API key was set. (#45928)

• Added
  Adds a check at startup to ensure requested models are accessible. (J#AI-80)
• Added
  Adds support for AWS Bedrock as a model provider. (J#AI-80)
• Added
  Adds support for OpenAI API-compatible model providers. (J#AI-80)
• Added
  Adds support for embedding models of any dimension size. (J#AI-80)
• Changed
  Improves the routing mechanism to better identify user intent. (#45716)
• Fixed
  Fixes an issue with suggested questions by relaxing the required question count. (#45715)

See the announcement post(opens in a new tab).

• Changed
  Release of 1.0.0.