Released: 26 Aug 2026
AIDB 7.6.0 introduces Agents and tools, a SQL-native framework for building in-database AI agents: an agent reasons over multiple steps — calling tools, observing their results, and repeating — instead of answering in one shot. Tools come from a unified catalog (aidb.tools) spanning a large built-in native tool set, your own custom SQL queries, and tools imported from an external MCP server. Agents support per-call budgets (tokens, reasoning iterations, wall-clock time), delegation to other agents, and per-user conversation visibility.
See Agents and Tools to get started.
AIDB 7.6.0 includes the following enhancements and bug fixes:
| Type | Description |
|---|---|
| Enhancement | Agents and tools: The new SQL-native agent framework and unified tools hub described above. See Agents and Tools. |
| Enhancement | Run any tool from SQL: The new aidb.run_tool() function runs a native, SQL, or MCP tool by name as the caller's own role, checks its arguments against the tool's declaration, and raises when an MCP tool reports an error. See aidb.run_tool. |
| Enhancement | Model cache budget: The new aidb.max_model_cache_size parameter caps the memory each process uses for loaded local models. AIDB evicts the least recently used models to stay under it and rejects a model larger than the limit. The default is a quarter of the host's RAM. See Model cache. |
| Enhancement | Native tool-calling model adapters: The new openai_responses and anthropic_messages providers (plus openai_responses_azure and anthropic_messages_azure/anthropic_messages_bedrock hosting variants) connect to OpenAI's Responses API and Anthropic's Messages API natively, emitting real tools/tool_choice request fields and parsing real tool-call responses, instead of simulating tool calls through prompt injection. See External model connections. |
| Enhancement | Model credentials from an environment variable: aidb.create_model() accepts a new credentials_env argument, letting credentials be read from a server-side environment variable at model-use time instead of passed inline as JSON. |
| Enhancement | Local reranking via llama.cpp: The new llamacpp_reranking provider reranks text locally from a GGUF file, supporting both cross-encoder models (for example, bge-reranker-v2-m3) and the Qwen3-Reranker decoder-only family — no external API required. See Local models. |
| Enhancement | Local OCR via llama.cpp: The new llamacpp_ocr provider runs OCR locally from a GGUF vision model, with a pre-registered default model (lightonocr-2-1b-Q8_0). See Local models. |
| Enhancement | Renamed decode_text to generate_text: aidb.decode_text() and aidb.decode_text_batch() are renamed to aidb.generate_text() and aidb.generate_text_batch(), matching encode_text()'s naming on the embedding side. The old names still work but are deprecated and will be removed in a future version. See Inference functions. |
| Enhancement | Semantic KB search across sources: The new aidb.semantic_kb_search() function fuses ranked vector search results from schema metadata and curated semantic aliases into a single result set using Reciprocal Rank Fusion. |
| Enhancement | Simplified semantic KB calls: kb_name (and name) are now optional on aidb.get_column_definitions(), aidb.get_metadata(), aidb.get_entity_definitions(), aidb.search_by_comment(), aidb.create_semantic_kb(), aidb.delete_semantic_kb(), aidb.refresh_semantic_kb(), aidb.update_semantic_kb_auto_processing(), and aidb.semantic_kb_stats() — when omitted, AIDB resolves the single existing semantic KB. |
| Change | Semantic aliases can belong to multiple KBs: aidb.create_semantic_alias() and related functions drop the model argument; an alias is now embedded once per knowledge base that owns the schema its SQL reads, instead of being bound to exactly one KB and model. |
| Enhancement | Pipeline metrics show last run time: aidb.get_pipeline_metrics() and the pipeline metrics views now include a last_run_completed timestamp. |
| Change | Leaked credential detection: aidb.create_model() now rejects a config that embeds api_key or basic_auth directly — pass them via credentials/credentials_env instead. The new aidb.audit_leaked_credentials() function reports any pre-existing models where credentials were detected embedded in their config. |
| Change | One error log row per failure: A pipeline's error log keeps one row per failing record, step, and operation. A repeated failure, including one logged by two executors at once, updates that row instead of adding a duplicate. The upgrade removes existing duplicates. See Pipeline error log. |
| Bug fix | Credentials with invalid header characters: An api_key or basic_auth value containing characters that aren't valid in an HTTP header (for example, a stray newline) now fails with a clear configuration error instead of crashing the backend. |