LEAI — Oracle Database Intelligence & Documentation Engine

LEAI (Lê - Aí in PT-BR) is an enterprise reverse engineering, impact analysis, and autonomous AI copilot engine for Oracle Database, specifically designed to power Retrieval-Augmented Generation (RAG), LLMs, and software engineers maintaining complex database ecosystems.

📑 Table of Contents / Índice

• 📌 What Is It?
• ⚙️ How It Works (Pipeline)
• 🤖 Autonomous Agent & Tool Calling Engine
• 🔗 Transparent Synonym Resolution
• ✂️ PL/SQL Semantic Compression
• 🚀 Quickstart: Using LEAI in Any Project
• 📖 CLI Command Reference
  • 1. Pipeline & Core Commands (leai, extract, annotate, compile, doc, generate)
  • 2. Impact Analysis & Lineage (leai trace)
  • 3. AI Copilot & Chat (leai ask, leai chat, leai models, leai enrich)
  • 4. Interactive Web Studio (leai serve)
  • 5. Specialized Subagents (leai agent)
  • 6. Autonomous Workflows (leai workflow)
  • 7. Business Rules & Canonical Glossary (leai rule)
  • 8. GitOps Version Control (leai git)
  • 9. S3 / SeaweedFS Distributed Storage (leai seaweed)
  • 10. Maintenance & Diagnostics (leai changes, leai init, leai doctor)
• 📁 Directory Structure
• 🧪 Automated Testing

📌 What Is It?

Enterprise Oracle databases accumulate years of business rules scattered across hundreds of tables, views, triggers, and massive PL/SQL packages (3,000 to 10,000+ lines of code).

Enabling developers or AI assistants to reliably understand such environments is challenging due to three main issues:

1. Token Inefficiency & Hallucinations: Sending entire monolithic packages into an LLM context is expensive, slow, and triggers attention degradation ("Lost in the Middle").
2. Hidden Dependencies: Altering a single column can silently break triggers, views, and procedures across multiple schemas.
3. Synonyms and Aliases: Stored procedures frequently access tables via private or public synonyms (PUBLIC SYNONYM), creating the false impression that referenced objects do not exist or belong elsewhere.

LEAI solves this by extracting the Oracle data dictionary, constructing a cross-schema dependency graph, and providing an autonomous multi-step reasoning agent with offline database tools.

⚙️ How It Works

LEAI operates via a 3-stage decoupled pipeline:

flowchart LR
    subgraph S1 [1. RAW JSON]
        DB[(Oracle Database)] -->|leai extract| RAW[Technical Snapshots<br/>./raw/*.json]
    end

    subgraph S2 [2. YAML Annotations]
        RAW -->|leai annotate| YAML[Business Annotations<br/>./annotations/*.yml]
        HUMAN[Human DBA / Engineer] -.-> YAML
        AI[LLM Auto-Enrich] -.-> YAML
    end

    subgraph S3 [3. DOCS & RAG]
        RAW & YAML -->|leai compile| DOCS[Markdown + Mermaid<br/>./docs/*.md]
        DOCS --> RAG[Vector Stores & LLMs]
    end

🤖 Autonomous Agent & Tool Calling Engine

When running leai chat or leai ask, the assistant uses an autonomous Tool-Calling Reasoning Loop (AgentExecutionEngine) with up to 10 iterations per turn. Instead of guessing or hallucinating, the model invokes specialized in-memory database tools:

🔗 Transparent Synonym Resolution

Stored procedures often access objects via PUBLIC SYNONYM or remote database links (@dblink). LEAI snapshots ALL_SYNONYMS and transparently dereferences every alias to its authentic physical entity, avoiding broken chains and LLM hallucinations.

✂️ PL/SQL Semantic Compression

For massive 10,000-line packages, LEAI extracts only the specific subprogram body requested while producing a lightweight signature skeleton of the rest of the package. This reduces token consumption by up to 95% while eliminating prompt distraction.

🚀 Quickstart: Using LEAI in Any Project

Step 1: Install LEAI

# Via pip
pip install leai

# Or via uv (recommended)
uv tool install leai
# Or in an existing project
uv add leai

Step 2: Initialize Configuration

# In English (default)
leai init

# Or in Brazilian Portuguese
leai init --lang pt-BR

Configure your leai.yml:

language: "en-US"                 # "en-US" (default) or "pt-BR"
update_check: true                # true (default) or false (disable with --no-update-check)
dsn: "oracle://${DB_USER}:${DB_PASS}@${DB_HOST}:1521/${DB_SERVICE}"

schemas:
  - HR
  - SALES

rawPath: "./raw"
annotationsPath: "./annotations"
docPath: "./docs"

ai:
  default_provider: "ollama"      # ollama, local, openai, gemini, anthropic, deepseek, qwen, kimi, grok
  temperature: 0.2
  timeout: 300.0
  max_history_turns: 15           # Chat history turns retained in memory
  max_agent_iterations: 10        # Maximum reasoning tool iterations per turn
  max_subagent_iterations: 5      # Maximum iterations for specialized subagents
  providers:
    ollama:
      base_url: "http://localhost:11434/v1"
      model: "qwen2.5-coder:latest"
      temperature: 0.1
      num_ctx: 32768              # Context window size for Ollama (prevents truncation on large DDLs)
      keep_alive: "1h"            # Keep model memory resident
    local:
      base_url: "http://localhost:1234/v1" # LM Studio, vLLM, LocalAI
      model: "qwen2.5"
      num_ctx: 32768
      max_tokens: 4096
    openai:
      api_key: "${OPENAI_API_KEY}"
      model: "gpt-4o-mini"
      max_tokens: 4096
    gemini:
      api_key: "${GEMINI_API_KEY}"
      model: "gemini-2.5-flash"

Step 3: Run the Full Pipeline

# 1. Run complete pipeline (extract + annotate + compile)
leai

# 2. Trace impact of modifying a table
leai trace EMPLOYEES --depth 2

# 3. Launch interactive terminal copilot
leai chat

# 4. Or launch the Web Studio in browser
leai serve

📖 CLI Command Reference

1. Pipeline & Core Commands

leai (or leai generate)

Executes the full automated pipeline: technical extraction, business annotation synchronization, and final Markdown compilation.

leai -s HR -t tables -t packages --depth 2 --rag-json

leai extract

Extracts Oracle data dictionary definitions into raw JSON snapshots.

# Incremental extraction of objects modified in the last 30 days
leai extract --days 30

leai annotate

Synchronizes YAML business annotation stubs under annotations/ without overwriting existing human notes.

leai doc <OBJECT>

Opens the in-terminal interactive documentation editor for a specific database entity.

leai doc EMPLOYEES
leai doc PKG_BILLING

leai compile

Recompiles the final Markdown documentation in docs/ with Mermaid.js diagrams.

2. Impact Analysis & Lineage (leai trace)

Generates multi-level upstream/downstream dependency trees, automated risk scores (LOW, MEDIUM, HIGH, CRITICAL), and Mermaid diagrams.

leai trace CONTRACTS_TB --depth 3 --offline --output ./dossier.md

3. AI Copilot & Chat

leai ask <QUESTION>

Answers one-off natural language queries about database structure and business rules.

leai ask "Which procedures update customer status to INACTIVE?" -p gemini

leai chat

Launches the interactive terminal copilot console with conversation memory, syntax highlighting, and live tool execution.

In-Session Slash Commands:

• /tune <sql>: Analyze query sargability, FTS risks, compound indexes, and get AI tuning proposals.
• /validate <sql>: Validate Oracle SQL dialect compliance and cross-check against schema catalog.
• /thoughts [on|off]: Toggle live reasoning/thought token streaming in the terminal.
• /workflow <name> <obj>: Execute autonomous workflows (reverse-procedure, impact-analysis, safe-refactor).
• /agent <role> <task>: Run specialized subagents (catalog_researcher, plsql_analyst, lineage_auditor, etc.).
• /copy [all|code|N]: Copy response or code block directly to OS clipboard.
• /doc [obj]: In-terminal YAML annotation & documentation editor.
• /rule [list|add|del|find]: Manage global domain glossary and canonical rules.
• /enrich [obj]: Auto-enrich business rules with LLM.
• /compile [obj]: Recompile Markdown docs (supports single object).
• /trace <obj>: Inline dependency & impact X-ray with Mermaid.
• /tables: List all catalog tables with column counts and primary keys.
• /schema [s]: Show full overview of schema objects.
• /changes [d]: Audit objects modified in last N days (Default: 7).
• /provider <p>: Switch AI provider dynamically (ollama, openai, gemini, anthropic, local, etc.).
• /models [p]: List available AI models returned by provider API.
• /model <p> [m]: Switch AI model dynamically.
• /audit [last|session|export]: Inspect AI tool call trace and latency.
• /tools: Quick viewer for last turn's tool execution inputs/outputs.
• /save [file.md]: Export current conversation transcript to Markdown.
• /clear: Clear conversation memory and reset screen.
• /exit, /quit: Exit interactive copilot session.

Smart Autocompletion Prefixes:

• @: Autocomplete catalog objects with icons (📋 TABLE, 📦 PACKAGE, ⚙️ PROCEDURE, 👁️ VIEW, ⚡ TRIGGER, 🔢 SEQUENCE, 🔗 SYNONYM).
• #: Autocomplete business rules and canonical terms from the domain glossary.
• /: Autocomplete slash commands and sub-arguments.

leai enrich

Invokes the LLM to inspect DDLs and draft automated business descriptions for undocumented entities.

leai enrich -o EMPLOYEES --overwrite -p gemini

leai models

Lists all configured AI providers, benchmarks network latency, and validates API keys.

4. Interactive Web Studio (leai serve)

Launches the visual LEAI Web Documentation & Annotation Studio daemon for in-browser collaborative annotation, instant Markdown compilation, and streaming AI copilot chat.

• Real-time SeaweedFS S3 Sync: Annotation edits in the browser (POST /api/annotations) are saved locally and synced directly to the S3 bucket in real time.
• Remote Fallback: Loads annotations directly from SeaweedFS S3 if not present on local disk.

leai serve --host 0.0.0.0 --port 9000

5. Specialized Subagents (leai agent)

Executes isolated technical personas with restricted, laser-focused database toolsets:

• leai agent list: Lists registered subagents.
• leai agent run <ROLE> <TASK>: Executes a subagent.

leai agent run plsql_analyst "Explain the interest calculation algorithm in PKG_BILLING and suggest index tuning"

6. Autonomous Workflows (leai workflow)

Multi-step orchestrated pipelines for high-risk engineering and reverse-engineering tasks:

• leai workflow list: Lists available workflows (impact-analysis, safe-refactor, reverse-procedure).
• leai workflow run <NAME> <TARGET>: Executes a workflow.

# Comprehensive impact dossier before modifying a table
leai workflow run impact CUSTOMERS_TB --output ./customers_impact.md

# Safe phased refactoring plan and DDL patch
leai workflow run refactor PKG_BILLING -p claude

# Decompile and specify business rules of a PL/SQL procedure with Mermaid diagram
leai workflow run reverse-procedure PRC_ATUALIZA_SALARIO --output ./specs.md

7. Business Rules & Canonical Glossary (leai rule)

Codifies domain concepts and canonical SQL predicates so AI copilots generate accurate queries:

• leai rule list: Lists codified business rules.
• leai rule add <TERM>: Adds a canonical rule.
• leai rule show <TERM>: Displays rule specifications.

leai rule add "ACTIVE_CUSTOMER" \
  --table "CUSTOMERS_TB" \
  --canonical-filter "RECORD_STATUS = 'A' AND IS_LOCKED = 0" \
  --definition "Customers eligible to place orders and receive billing invoices" \
  --tags "sales,compliance"

8. GitOps Version Control (leai git)

Treats documentation as first-class code (Docs-as-Code):

• leai git status [--fetch]: Inspects repository status across tracked documentation paths.
• leai git pull: Pulls latest annotations from remote Git repository.
• leai git sync [-m "message"]: Stages, commits, and pushes modified annotations and docs.

leai git sync -m "docs(billing): update tax calculation business rules"

9. S3 / SeaweedFS Distributed Storage (leai seaweed)

Collaborative metadata persistence using S3-compatible Object Storage:

• leai seaweed status: Verifies S3 connection and bucket health.
• leai seaweed push: Uploads local snapshots to remote S3 bucket.
• leai seaweed pull: Downloads snapshots from remote S3 bucket.
• Web Studio Integration: Edits made in Web Studio (/serve) sync directly to SeaweedFS in real time.
• Lifecycle Rules: Compatible with standard S3 lifecycle configurations (NoncurrentVersionExpiration on annotations/) to purge old non-current versions automatically.

10. Maintenance & Diagnostics

• leai changes: Audits database objects modified in the last N days via Oracle's LAST_DDL_TIME (-d, --days, -u, --user).
• leai doctor (or check): Pre-flight verification of Oracle connectivity, catalog permissions, pipeline directories, and AI credentials.
• leai init: Generates a starter leai.yml (-L, --lang [en-US|pt-BR], -f, --force, -o, --output).

📁 Directory Structure

my_project/
├── leai.yml                  <-- Master configuration file
├── raw/                      <-- Raw technical JSON snapshots extracted from Oracle
│   └── HR.json
├── annotations/              <-- Human & AI business rules in YAML (non-destructive)
│   └── HR.yml
└── docs/                     <-- Final compiled Markdown documents for RAG and humans
    └── HR/
        ├── INDEX.md          <-- Master navigation index
        ├── tables/           <-- Tables with Mermaid diagrams and YAML frontmatter
        └── code_objects/     <-- Procedures, packages, functions, views

🧪 Automated Testing

To run the complete automated test suite:

# Run unit tests with test coverage reporting
uv run coverage run -m unittest discover tests
uv run coverage report -m

# Run code linter
uv run ruff check .