apex-rag 0.1.3

A production-grade, local-first Agentic RAG library using structural document navigation.

ApexRAG

Production-grade, local-first Agentic RAG Library. Replaces vector similarity search with structural, agentic navigation of documents.

---

🧠 The Core Idea

Traditional RAG embeds text into vectors and finds the "closest" chunks. This creates retrieval hallucinations — the model returns semantically-similar-but-wrong content because it has no understanding of document structure.

ApexRAG takes a fundamentally different approach:

1. Parse the document into a structural tree (based on headings) and extract page numbers.
2. Synthesize a 30-word Semantic Map for every node using a local LLM.
3. Navigate the tree with an LLM agent that reads summaries and decides which branch to enter — trying multiple candidates if necessary.
4. Verify the exact leaf node answers the query via a strict secondary LLM check (99.999% accuracy).
5. Return the exact leaf node content — not a blended, hallucinated average.

Query: "What were Q3 revenues?"
         │
    Root (Annual Report)
    ├── Chapter 1: Executive Summary  ← LLM: "Not here"
    └── Chapter 2: Revenue Analysis   ← LLM: "Enter this"
        ├── Q1 Revenue                ← LLM: "Not Q3"
        ├── Q2 Revenue                ← LLM: "Not Q3"
        └── Q3 Revenue                ← LLM: "This is it!" → Return content

---

📁 Project Structure

apex_rag/
├── src/
│   ├── __init__.py       # Public API exports
│   ├── api.py            # FastAPI App & UI dashboard
│   ├── client.py         # Thread-safe user-facing ApexIndex class
│   ├── ingestion.py      # Document parsing & tree synthesis
│   ├── navigation.py     # Recursive LLM navigation agent
│   ├── storage.py        # SQLAlchemy async ORM & PageIndexEntry
│   └── utils.py          # ReasoningTrace, retry decorator, helpers
├── tests/
│   ├── test_tree.py      # Parser & storage unit tests
│   └── test_search.py    # Navigation agent unit tests (no Ollama needed)
├── examples/
│   └── basic_usage.py    # End-to-end demo
├── pyproject.toml
└── docker-compose.yml

---

⚡ Quick Start

1. Install

# Clone and set up
cd ApexRAG
pip install -e ".[dev]"

2. Start Ollama

ollama serve
ollama pull llama3.1   # or phi3, mistral, etc.

3. Ingest & Query

import asyncio
from src.client import ApexIndex

async def main():
    async with await ApexIndex.create(
        db_url="sqlite+aiosqlite:///apex.db",
        model="llama3.1",
    ) as index:
        # Ingest a PDF
        doc_id = await index.ingest("path/to/your/report.pdf")

        # Query it
        result = await index.query(
            "What are the Q3 revenue figures?",
            doc_id,
        )

        if result:
            print(result.content)
            print(f"Found at path: {result.path}")
            print(f"Navigation trace: {result.trace}")

asyncio.run(main())

4. Start the FastAPI Server & Visual Index Dashboard

uvicorn src.api:app --reload

Open your browser to:

• Dashboard: http://localhost:8000
• API Docs: http://localhost:8000/docs

From the dashboard, you can click on an ingested document to view its full structural tree and its book-style alphabetical page index!

---

🏗️ Architecture Deep Dive

Ingestion Engine (ingestion.py)

Step	Description
Convert	markitdown or docling converts PDF/DOCX → Markdown
Parse	Regex walks ATX headings (#, ##, ###) to build ParsedSection tree
Persist	Nodes written to DB with LTree-style path (1.2.3)
Synthesize	Ollama generates 30-word summaries in parallel (bounded by semaphore)

Storage Layer (storage.py)

DocumentNode table:
  id          BIGINT PRIMARY KEY
  doc_id      VARCHAR(255)       -- logical document identifier
  parent_id   BIGINT FK (self)   -- NULL for root nodes
  path        VARCHAR(512)       -- "1.2.3" LTree-style
  title       VARCHAR(512)       -- section heading
  summary     TEXT               -- 30-word Semantic Map
  content     TEXT               -- leaf content (NULL for intermediate)
  metadata    TEXT (JSON)        -- page numbers, char count, source file
  depth       INTEGER            -- nesting level (0 = root)
  position    INTEGER            -- sibling order
  created_at  TIMESTAMP

Supports both sqlite+aiosqlite:// (local) and postgresql+asyncpg:// (production).

Navigation Agent (navigation.py)

find(query, doc_id)
  └── _navigate(current_node)
        ├── [Leaf?] → return content immediately
        ├── fetch children
        ├── _ask_llm(query, child_summaries)
        │     └── "Which child ID contains the answer?"
        ├── [ID returned] → recurse into chosen child
        │     └── [child returns None] → try siblings
        └── [NONE returned] → backtrack to parent

LLM Response Parsing is robust — 4-tier fallback:

1. Strict json.loads()
2. Regex extraction from prose-wrapped JSON
3. Explicit "NONE" keyword detection
4. Heuristic: scan for any valid child ID number in the response

High Accuracy (99.999%) Verification: At the leaf level, a second LLM prompt strictly verifies if the leaf content answers the query. If it fails, the agent backtracks and explores the fallback candidates (second-best choices) up the tree.

Reasoning Trace (utils.py)

Every navigation decision is printed with color-coded indicators:

━━━ ApexRAG Navigation Start ━━━
Query : What are the Q3 revenue figures?
Root  : node_id=1

  ↳ ENTER node=1 path=1
    Covers the full annual financial report for 2024…
    ⟳ EXPLORE node=1 → evaluating 2 child summaries
    ✔ AGENT → node=3  reason: Revenue Analysis contains quarterly breakdown
    ↳ ENTER node=3 path=1.2
      ⟳ EXPLORE node=3 → evaluating 4 child summaries
      ✔ AGENT → node=6  reason: Q3 Revenue section is exactly what's needed
      ★ LEAF REACHED node=6
        preview: Q3 revenue was $165M. Growth slowed slightly…

━━━ Navigation Complete ━━━  result=SUCCESS  elapsed=3.41s

---

🧪 Testing

# Run all tests (no Ollama required)
pytest

# With coverage
pytest --cov=src --cov-report=term-missing

# Specific test file
pytest tests/test_search.py -v

Tests use an in-memory SQLite database and mock LLM responses — zero external dependencies.

---

🐳 Production Deployment

# Copy and edit environment
cp .env.example .env

# Start everything (Ollama + PostgreSQL + API)
docker-compose up -d

# Pull the model inside the Ollama container
docker exec apex_ollama ollama pull llama3.1

Environment variables:

Variable	Default	Description
APEX_DB_URL	sqlite+aiosqlite:///apex.db	SQLAlchemy async DB URL
APEX_OLLAMA_HOST	http://localhost:11434	Ollama server URL
APEX_MODEL	llama3.1	Ollama model for navigation
APEX_LOG_LEVEL	INFO	Logging verbosity

---

🔧 Configuration Reference

await ApexIndex.create(
    db_url="postgresql+asyncpg://user:pass@host/db",  # Production DB
    ollama_host="http://localhost:11434",
    model="llama3.1",              # Navigation model
    summariser_model="phi3",       # Cheaper model for ingestion summaries
    max_concurrent_summaries=8,    # Parallelism (tune to your GPU VRAM)
    parser_backend="markitdown",   # "markitdown" | "docling" | "plaintext"
    trace_enabled=True,            # Color-coded console output
    db_echo=False,                 # SQL query logging
)

---

📋 Roadmap

• FastAPI REST API wrapper (/documents/ingest/file, /query, /documents)
• Book-style Page Index and Visual tree dashboard
• Unlimited navigation depth with backtrack and verification
• Streaming query responses via SSE
• Multi-document cross-reference queries
• Alembic migrations for schema versioning
• Support for docling table extraction (structured data cells as leaf nodes)

---

📄 License

MIT License — see LICENSE.