Skip to content

How It Works ​

The problem ​

AI agents need to find information in documents. Traditional approaches use vector databases — they split documents into chunks, embed them, and do similarity search. This loses document structure, mixes sections, and requires maintaining an embedding pipeline.

The approach ​

Logidoc builds a table of contents with summaries for each section. An agent reads the TOC, reasons about which sections are relevant, and retrieves only what it needs. No vectors, no embeddings.

Indexation pipeline ​

When you upload and index a document, this happens:

Upload (stored in MongoDB)
  │
  ▼
Parse: extract text page by page
  │  PDF → pdftotext (poppler)
  │  DOCX/PPTX/HTML → pandoc → markdown
  │
  ▼
Detect structure
  │  ┌─ TOC with page numbers? → extract directly, stop
  │  ├─ TOC without page numbers? → search for titles in text
  │  └─ No TOC? → scan in chunks of 10 pages
  │
  ▼
Calibrate page numbers
  │  Logical pages (from TOC) → physical PDF pages
  │  Detects title page / preface offset automatically
  │
  ▼
Build tree + subdivide large sections
  │  Flat sections → hierarchical tree by level
  │  Sections > 20 pages → LLM splits into sub-sections
  │
  ▼
Enrich content (optional)
  │  Tables → heuristic detection + VLM fallback
  │  Images → pdfimages extraction + VLM description
  │
  ▼
Fill text from pages → save to MongoDB

Two models ​

Logidoc supports two separate models for different tasks:

  • Main model (LLM_*) — structure detection, TOC extraction, chunking, subdivision. Text-only, fast, cheap.
  • Vision model (VISION_*) — table VLM fallback, image description. Needs vision capability.

If no vision model is configured, the main model is used for everything.

env
LLM_PROVIDER=anthropic
LLM_MODEL=claude-haiku-4-5-20251001

VISION_PROVIDER=openai
VISION_MODEL=gpt-4o-mini

Retrieval ​

Agents never see the full document. They interact through 4 tools:

  1. list_documents → see what's available
  2. search → keyword search across all TOCs (returns matching sections)
  3. get_toc → browse the structure (titles + summaries + page numbers, no text)
  4. get_sections → get full text of specific sections by ID

A typical flow:

Agent: search("SSE streaming")
  → 3 matching sections across 2 documents

Agent: get_sections(doc_id, "sse-streaming,sse-implementation")
  → full text of those 2 sections

Agent: answers the user's question with citations

Released under the MIT License.