1.18.0

2026-06-19

BREAKING CHANGES

• PromptRegistry rewritten on Jinja2: PromptLanguage/ModelName enums and old register/value/instruction/tag API removed; no longer a singleton.
• Local embedding routing prefix changed from local/ to local: in EMBEDDING_ENGINE_MODEL.
• gemini-3.1-flash-lite-preview replaced by GA gemini-3.1-flash-lite (preview shut down).

Added

• Local embedding engine backed by SentenceTransformer, with L2 normalization and dimension truncation.
• Jinja2 PromptRegistry: load_from_folder, manifest.toml, render, register_template, has_template.
• normalize_embeddings option on VectorDB (default off).
• gemini-3.5-flash added to supported reasoning models.
• Parallel Search v1 params: location, fetch_policy, client_model, session_id, after_date, mcp_servers, previous_interaction_id.
• client_timeout/client_max_retries on DynamicEngine and the OpenAI search engine.
• Prompt Registry guide in docs.

Changed

• Parallel Search migrated from beta to v1 API (client.search/client.extract, advanced_settings).
• Default Parallel search mode changed from one-shot to advanced.
• parallel-web requirement raised to >=1.0.1 for scrape and search extras.
• VectorDB selects local vs cloud embeddings by model name instead of API key.

Fixed

• Hung web_search calls now fail fast via configurable socket timeout.
• DynamicEngine forwards client_timeout/client_max_retries to neurosymbolic engines.
• Hard sys.exit(1) on unconfigured neurosymbolic engine replaced with warnings.
• Removed erroneous TEXT_TO_SPEECH API-key auto-migration.
• Local embedding bugs: id() routing, model priority, and scalar-vector shape handling.
• Optional engine constructors skip SDK init when inactive, eliminating noisy registration warnings.
• Duplicate SentenceTransformer imports removed.

Security

• Dropped trust_remote_code=True from local embeddings, removing arbitrary code execution risk.

Maintenance

• Tests added for local embedding, Parallel Search citations/geo-targeting, and DynamicEngine timeout forwarding.