malimgraph 0.1.2

Transform PDF documents into structured knowledge graphs with citation provenance

MalimGraph

███╗   ███╗ █████╗ ██╗     ██╗███╗   ███╗ ██████╗ ██████╗  █████╗ ██████╗ ██╗  ██╗
████╗ ████║██╔══██╗██║     ██║████╗ ████║██╔════╝ ██╔══██╗██╔══██╗██╔══██╗██║  ██║
██╔████╔██║███████║██║     ██║██╔████╔██║██║  ███╗██████╔╝███████║██████╔╝███████║
██║╚██╔╝██║██╔══██║██║     ██║██║╚██╔╝██║██║   ██║██╔══██╗██╔══██║██╔═══╝ ██╔══██║
██║ ╚═╝ ██║██║  ██║███████╗██║██║ ╚═╝ ██║╚██████╔╝██║  ██║██║  ██║██║     ██║  ██║
╚═╝     ╚═╝╚═╝  ╚═╝╚══════╝╚═╝╚═╝     ╚═╝ ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝     ╚═╝  ╚═╝

From documents to knowledge graphs.

Transform PDF documents into structured knowledge graphs with full citation provenance. Every entity and relationship traces back to the exact PDF page and verbatim text that supports it.

---

Features

Tool	Description
extract_knowledge_graph	Hybrid rule + LLM extraction → entities, relationships, citations
chunk_document	Token-aware overlapping chunks with heading context for RAG
render_document_html	Structured HTML with page anchors, entity annotations, TOC + search
manage_graph_db	Load, query, and manage graphs in Neo4j or PostgreSQL (Apache AGE)
embed_and_store_chunks	Embed chunks into PostgreSQL pgvector (OpenAI / Voyage / local)

Three ways to use:

• MCP Server — connect to Claude Desktop, Claude Code, or claude.ai
• CLI — malimgraph extract, chunk, render, db, vector
• Claude Skills — 5 installable .skill packages for claude.ai

---

Quick Start

Claude Code / Claude Desktop (no API key needed)

pip install malimgraph
claude mcp add malimgraph -- malimgraph-plugin

Restart Claude Code, then just ask:

"Extract a knowledge graph from report.pdf and save to ./output/"

Claude reads the PDF, extracts entities using its own intelligence, and saves the graph. No ANTHROPIC_API_KEY required.

CLI (standalone, requires API key)

pip install malimgraph
export ANTHROPIC_API_KEY=sk-ant-...

malimgraph extract --input report.pdf --output ./output/ --format all
malimgraph chunk --input report.pdf --output ./chunks/
malimgraph render --input report.pdf --output document.html

---

How It Works

PDF
 │
 ▼
pdf_reader.py ──────────────────────────────────────────────┐
 │  (PyMuPDF: text, headings, tables, page structure)       │
 ├──────────────────────────────────┐                        │
 ▼                                  ▼                        ▼
rule_extractor.py              llm_extractor.py          chunker.py
 │ (regex: dates, amounts,      │ (Anthropic API:         │ (sliding window
 │  emails, legal refs,         │  semantic entities,     │  with heading
 │  section numbers)            │  relationships,         │  context)
 │                              │  source_text required)  │
 └──────────────┬───────────────┘                         │
                ▼                                          ▼
          graph_builder.py                          embedder.py
           │ (merge + dedup:                        │ (OpenAI / Voyage /
           │  hybrid method,                        │  local sentence-
           │  citation accumulation,                │  transformers)
           │  stable IDs)                           │
           ▼                                        ▼
     knowledge_graph.json                    vector_client.py
           │                                 (pgvector: HNSW index,
     ┌─────┴──────┐                           cosine similarity search)
     ▼             ▼
 cypher.py     age_sql.py
 (.cypher)      (.sql)

---

Three Ways to Use

Claude Code Plugin (recommended — no API key)

claude mcp add malimgraph -- malimgraph-plugin

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "malimgraph": {
      "command": "malimgraph-plugin"
    }
  }
}

Claude uses its own subscription to extract entities — no ANTHROPIC_API_KEY needed. See docs/claude-code-plugin.md for full details.

MCP Server (standalone / HTTP)

# stdio — Claude Desktop / Claude Code (with API key for LLM extraction)
malimgraph serve

# HTTP — remote connections / claude.ai
malimgraph serve --transport http --port 8080

CLI

# Extract knowledge graph from PDF
malimgraph extract \
  --input report.pdf \
  --output ./output/ \
  --entity-types auto \
  --format all \
  --graph-name my_graph

# Chunk for embeddings
malimgraph chunk \
  --input report.pdf \
  --output ./chunks/ \
  --chunk-size 512 \
  --overlap 64 \
  --format json

# Embed chunks into PostgreSQL pgvector
malimgraph vector load \
  --input ./chunks/chunks.json \
  --uri "postgresql://user:pass@localhost:5432/mydb" \
  --provider openai \
  --table document_chunks

# Semantic search over embedded chunks
malimgraph vector search \
  --query "What are the financial risks?" \
  --uri "postgresql://user:pass@localhost:5432/mydb" \
  --top-k 5

# Render as browsable HTML
malimgraph render \
  --input report.pdf \
  --output document.html \
  --knowledge-graph ./output/knowledge_graph.json

# Load into Neo4j
malimgraph db load \
  --input ./output/knowledge_graph.json \
  --target neo4j \
  --uri bolt://localhost:7687 \
  --user neo4j \
  --password secret

# Query the graph
malimgraph db query \
  --target neo4j \
  --uri bolt://localhost:7687 \
  --query "MATCH (n:Organization) RETURN n.label, n.source_pages LIMIT 10"

Claude Skills

Download .skill files from GitHub Releases and install in claude.ai → Settings → Skills.

Skill	Trigger phrases
pdf-to-knowledge-graph	"knowledge graph", "extract entities", "PDF to Cypher"
pdf-to-chunks	"chunk document", "split for embeddings", "RAG chunks"
document-to-html	"convert PDF to HTML", "render document", "make PDF browsable"
graph-db-admin	"load into Neo4j", "Cypher query", "graph statistics"
chunks-to-pgvector	"store in pgvector", "embed into PostgreSQL", "semantic search", "RAG with PostgreSQL"

---

Installation

# Core (knowledge graph + chunking + HTML)
pip install malimgraph

# With Neo4j support
pip install "malimgraph[neo4j]"

# With Apache AGE support
pip install "malimgraph[age]"

# With pgvector + OpenAI embeddings
pip install "malimgraph[pgvector,openai]"

# With pgvector + Voyage AI embeddings
pip install "malimgraph[pgvector,voyage]"

# With local embeddings (no API key needed)
pip install "malimgraph[pgvector,local]"

# Everything
pip install "malimgraph[all]"

Environment Variables

ANTHROPIC_API_KEY=sk-ant-...      # Required for LLM extraction
OPENAI_API_KEY=sk-...             # Required for OpenAI embeddings
VOYAGE_API_KEY=pa-...             # Required for Voyage AI embeddings
PGVECTOR_URI=postgresql://...     # PostgreSQL connection for pgvector
NEO4J_URI=bolt://localhost:7687   # Neo4j connection
NEO4J_USER=neo4j
NEO4J_PASSWORD=yourpassword
AGE_CONNECTION_URI=host=...       # Apache AGE connection

---

Output Schema — knowledge_graph.json

Every entity and relationship carries full citation provenance:

Field	Type	Description
id	string	Stable hash ID: e_ + MD5(type:label)[:8]
label	string	Canonical entity name
type	string	Entity type (Organization, Person, Date, …)
extraction_method	enum	rule / llm / hybrid
confidence	enum	high / medium / low
source_pages	int[]	PDF page numbers where found
source_text	string	Primary verbatim supporting quote
source_chunk_id	string	Processing chunk ID
citations[]	object[]	All supporting quotes with page refs
citation_count	int	Stored as property in graph DBs

---

pgvector — Semantic Search Schema

Chunks are stored with embeddings in PostgreSQL, enabling semantic search:

-- Find chunks most similar to a query
SELECT chunk_text, source_file, page_numbers, heading_context,
       1 - (embedding <=> '[...]'::vector) AS score
FROM document_chunks
ORDER BY embedding <=> '[...]'::vector
LIMIT 10;

-- Filter by document
SELECT * FROM document_chunks
WHERE document_id = 'annual_report_2024'
ORDER BY embedding <=> '[...]'::vector LIMIT 5;

Supported embedding providers:

Provider	Default model	Dimension	API key
openai	text-embedding-3-small	1536-d	OPENAI_API_KEY
voyage	voyage-3-large	1024-d	VOYAGE_API_KEY
local	all-MiniLM-L6-v2	384-d	none (CPU)

---

Database Setup

Neo4j

docker run -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/yourpassword neo4j:latest

Apache AGE (PostgreSQL)

docker run -p 5432:5432 -e POSTGRES_PASSWORD=secret apache/age:latest

pgvector (PostgreSQL)

docker run -p 5432:5432 -e POSTGRES_PASSWORD=secret pgvector/pgvector:pg17

See docs/database-setup.md for full guides.

---

Contributing

1. Fork the repo
2. Create a feature branch: git checkout -b feature/my-feature
3. Install dev deps: pip install -e ".[dev]"
4. Run tests: make test
5. Lint: make lint
6. Submit a PR

---

Credits

Built by Malim AI Labs — AI-powered knowledge infrastructure for Southeast Asia.

Malim AI Labs Social Enterprise (003827047-U) · Kuala Lumpur, Malaysia

---

License

MIT — see LICENSE