Autonomous Agent & In-Memory Tools

LEAI features an autonomous reasoning engine based on the ReAct (Reason + Act) paradigm via its AgentExecutionEngine. Instead of guessing database structures or overwhelming prompt windows, the agent actively inspects catalog metadata in real time with thought/reasoning token streaming and in-memory tool calling.

🔄 How the Execution Loop Operates

sequenceDiagram
    autonumber
    actor Dev as User / Engineer
    participant Agent as LEAI Agent Loop
    participant Tools as In-Memory Database Tools
    participant LLM as Language Model (LLM)

    Dev->>Agent: "Which procedures update customer status to INACTIVE?"
    Agent->>LLM: Evaluate intent (real-time thought/reasoning token streaming)
    LLM-->>Agent: Invoke search_catalog(query='INACTIVE')
    Agent->>Tools: search_catalog
    Tools-->>Agent: Matches: PROC_DEACTIVATE_CUST, TRG_STATUS_AUDIT
    Agent->>LLM: Evaluate tool output
    LLM-->>Agent: Invoke get_subprogram_source(subprogram_name='PROC_DEACTIVATE_CUST')
    Agent->>Tools: get_subprogram_source (with semantic compression)
    Tools-->>Agent: Returns authentic procedure body
    Agent->>LLM: Synthesize final answer (token-by-token streaming)
    LLM-->>Dev: Verified answer with authentic code, rules, and dependencies

The agent runs with a configurable safety guard of up to 10 reasoning cycles per turn to prevent infinite loops.

⚡ Real-Time Streaming & Thought/Reasoning Tokens

LEAI supports synchronous and asynchronous streaming with Tool Calling (stream_chat_with_tools) across all major LLM providers: * DeepSeek-R1 / Qwen 2.5 / Ollama: Captures and emits <think> / reasoning_content blocks via the on_thought event. * Gemini (Google AI): Extracts native thought: true and thoughtSignature blocks via Server-Sent Events (SSE). * Anthropic Claude: Processes native thinking_delta / thinking blocks and streams incremental tool arguments (input_json_delta). * Web Studio & CLI: Live reasoning traces appear dynamically in terminal and web interfaces (leai serve) without blind spinner pauses.

🛠️ Available Agent Tools

Tool	Parameters	Purpose
`search_catalog`	`query`, `object_types`, `schema`	Fast text and regex search across tables, views, procedures, packages, functions, and synonyms.
`get_table_schema`	`table_name`, `schema`	Comprehensive DDL inspection: data types, nullable constraints, PKs, FKs, unique keys, check constraints, and indexes.
`get_subprogram_source`	`package_name`, `subprogram_name`	Surgical extraction of standalone procedure/function or subprogram inside a package (with semantic compression).
`grep_plsql_code`	`pattern`, `object_name`, `schema`	Fast regex search across all stored PL/SQL bodies without reading entire packages.
`trace_object_lineage`	`object_name`, `schema`, `depth`, `direction`	Multi-level upstream/downstream dependency graph with automated refactoring risk score (`LOW` to `CRITICAL`).
`explain_and_tune_sql`	`sql_query`, `detailed`	Evaluates sargability (`TRUNC`, `NVL`, `UPPER`), Full Table Scan (FTS) risks, `NOT IN (SELECT ...)` NULL pitfalls, compound index ordering, and AI query rewrites.
`validate_oracle_sql`	`sql_query`, `target_schema`	Validates Oracle dialect compliance, blocks non-Oracle constructs (`LIMIT`, `BOOLEAN`, `ILIKE`, `IFNULL`, `+` concat), and checks against schema catalog.
`lookup_business_term`	`query`, `tag`	Searches domain glossary for canonical business definitions, calculation rules, and canonical SQL predicates.
`estimate_query_cost`	`sql_query`	Estimates query complexity and relative execution cost.
`query_schema_metadata`	`schema_name`	Retrieves aggregated object counts and schema summaries.

💡 Benefits of Offline Tool-Calling

Zero Access to Live Data: Tools query the extracted metadata snapshot, ensuring data privacy and enterprise compliance (GDPR/LGPD/SOC2).
Eliminates Hallucinations: The model confirms facts by reading verified DDLs and code before forming its response.
Token Efficiency: Only relevant subprograms and table definitions enter the prompt window.