rapidGraph 0.2.0

Open-domain text-to-graph extractor with entities, relations, schema inference, and Neo4j export.

rapidGraph

rapidGraph is a local-first, open-domain text-to-graph extractor for arbitrary text. It reads inline text or one or more .txt files and produces a structured graph-oriented JSON payload with:

• entities
• relations
• potential_schema
• expanded_schema
• documents
• chunks
• relation_support
• meta

The package is designed for:

• entity and relation extraction across general, technical, scientific, and mixed-domain text
• CPU-friendly local execution
• provenance-aware graph construction
• future GraphRAG / RAG workflows
• optional Neo4j ingestion
• optional Neo4j vector GraphRAG question answering with Ollama

The public package name is rapidGraph, the import package is rapidgraph, and the installed CLI command is rapidgraph.

Table of Contents

• What rapidGraph Does
• Key Capabilities
• Installation
• Quick Start
• How the Pipeline Works
• Execution Modes
• Input Model
• Output Model
• CLI Reference
• Recommended Flag Combinations
• Neo4j Export Model
• GraphRAG Ask Mode
• Python Library Usage
• Performance and Practical Notes
• Troubleshooting
• Development
• Publishing
• License

What rapidGraph Does

At a high level, rapidGraph takes arbitrary text and turns it into a graph-friendly representation.

The pipeline:

1. normalizes input text
2. splits text into chunked spans
3. extracts entity candidates
4. extracts relation candidates
5. canonicalizes duplicate or near-duplicate mentions
6. links relation endpoints to canonical entities
7. infers schema patterns from the accepted graph edges
8. stores document and chunk provenance so every entity mention and relation can be traced back to source text

The extractor is open-domain and best-effort. It does not rely on a fixed business-only ontology. If typing confidence is weak, it keeps entities as Unknown rather than discarding them.

Key Capabilities

• Open-domain entity extraction
• Open-domain relation extraction
• Schema inference from extracted graph edges
• Provenance-aware output using documents, chunks, and relation_support
• Multi-file corpus ingestion in one run
• Two entity canonicalization scopes:
  • document
  • corpus
• Three execution modes:
  • fast
  • balanced
  • quality
• Optional embedding-assisted entity merging and relation endpoint linking
• Optional Neo4j export
• Optional GraphRAG query layer over Neo4j vector indexes and Ollama
• Backward-compatible potential_schema plus richer expanded_schema

Installation

Install from PyPI

pip install rapidGraph

Install with optional extras

Neo4j support:

pip install "rapidGraph[neo4j]"

Embedding-assisted linking:

pip install "rapidGraph[embeddings]"

GraphRAG query support:

pip install "rapidGraph[graphrag]"

Development tooling:

pip install "rapidGraph[dev]"

Everything:

pip install "rapidGraph[neo4j,embeddings,graphrag,dev]"

Install from source

pip install .

Or with extras:

pip install ".[neo4j,embeddings,graphrag,dev]"

Quick Start

Show CLI help:

rapidgraph --help

Extract from inline text:

rapidgraph --text "Google is based in California." --pretty

Extract from a file:

rapidgraph --input input.txt --pretty

Extract from multiple files:

rapidgraph --input input.txt input2.txt --pretty

Write JSON to a file:

rapidgraph --input input.txt --output graph.json --pretty

Export to Neo4j with chunk embeddings and a vector index for GraphRAG:

rapidgraph \
  --input input.txt \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --neo4j-embed-chunks \
  --neo4j-create-vector-index

Ask the graph using Ollama:

rapidgraph ask \
  --question "What does the text say about attention?" \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --ollama-model llama3.2 \
  --pretty

The repo-root compatibility shim also works:

python extract_graph.py --input input.txt --pretty

How the Pipeline Works

1. Text normalization

The input is normalized for whitespace and line ending consistency before extraction begins.

2. Chunking

The extractor splits text into chunks before model inference. Chunking exists because relation and entity models work better on bounded spans than on arbitrarily long documents.

Two chunking strategies are available:

• paragraph
  • default
  • respects paragraph and block boundaries first
  • better for preserving local structure
• sentence
  • simpler sentence packing
  • useful for experimentation or tighter chunk control

Optional overlap preserves context across chunk boundaries.

3. Entity extraction

Entities are primarily extracted with GLiNER, with heuristic fallback and supplemental heuristics used where useful.

4. Relation extraction

Relations come from a combination of:

• heuristic relation extraction
• context-based relation patterns
• optional REBEL relation extraction

The amount of REBEL usage depends on --mode.

5. Canonicalization

Mentions are merged into canonical entities using:

• normalized string matching
• fuzzy matching
• optional embedding-assisted rescue for borderline cases

6. Relation linking

Relation endpoints are linked back to canonical entity IDs using:

• exact and local mention-aware matching first
• fuzzy matching second
• optional embedding-assisted rescue last

7. Schema generation

Two schema views are produced:

• potential_schema
  • strict compatibility view
  • grouped by (source_type, relation, target_type)
• expanded_schema
  • richer view using more refined type groupings
  • keeps more semantic detail

8. Provenance capture

Each mention and accepted relation can be traced to:

• a document
• one or more chunks
• representative evidence text

This is what makes the model usable later for retrieval or graph-backed answer generation.

Execution Modes

rapidGraph supports three runtime modes.

fast

Best for:

• CPU-only quick passes
• rapid iteration
• rough graph drafts

Behavior:

• uses GLiNER plus heuristics
• does not run REBEL
• lowest startup cost
• lowest relation recall of the three modes

Example:

rapidgraph --input input.txt --mode fast --pretty

balanced

This is the default mode.

Best for:

• most local CPU runs
• practical relation quality without paying the full REBEL cost

Behavior:

• runs heuristic relations everywhere
• runs REBEL only on shortlisted high-value spans
• usually the best speed/quality tradeoff

Example:

rapidgraph --input input.txt --mode balanced --pretty

quality

Best for:

• slower offline analysis
• smaller corpora
• maximum relation recall

Behavior:

• runs REBEL across all chunks
• highest model cost
• typically the slowest mode

Example:

rapidgraph --input input.txt --mode quality --pretty

Input Model

The CLI accepts exactly one of:

• --text "..." for inline text
• --input file1.txt [file2.txt ...] for file input

--text and --input are mutually exclusive.

Multi-file ingestion produces a single combined JSON result with multiple documents and chunks.

Output Model

The extractor returns a single JSON object.

entities

Each entity contains:

• id
• text
• canonical
• type
• confidence
• mentions

Each mention contains:

• text
• start
• end
• chunk_index
• document_id
• chunk_id

relations

Each relation contains:

• source_id
• target_id
• relation
• confidence
• evidence
• chunk_ids
• document_ids

potential_schema

Strict schema aggregation. This preserves backward compatibility and groups edges by:

• source_type
• relation
• target_type

expanded_schema

Richer schema aggregation that retains more type detail and gives a broader schema view than potential_schema.

documents

One row per input document:

• id
• source
• title
• text_hash
• char_count

chunks

One row per extraction chunk:

• id
• document_id
• index
• text
• start
• end
• block_index
• overlap_sentences

relation_support

One row per final accepted relation edge with merged provenance:

• source_id
• relation
• target_id
• chunk_ids
• document_ids
• evidence

meta

Contains execution metadata such as:

• model names
• thresholds
• chunk count
• elapsed time
• mode
• relation backend strategy
• REBEL usage counts
• embedding usage counts
• warning list
• fallback indicator

Example output shape

{
  "entities": [
    {
      "id": "E1",
      "text": "Google",
      "canonical": "Google",
      "type": "Organization",
      "confidence": 0.91,
      "mentions": [
        {
          "text": "Google",
          "start": 0,
          "end": 6,
          "chunk_index": 0,
          "document_id": "D1",
          "chunk_id": "D1:C0"
        }
      ]
    }
  ],
  "relations": [
    {
      "source_id": "E1",
      "target_id": "E2",
      "relation": "IS_BASED_IN",
      "confidence": 0.78,
      "evidence": "Google is based in California.",
      "chunk_ids": ["D1:C0"],
      "document_ids": ["D1"]
    }
  ],
  "potential_schema": [],
  "expanded_schema": [],
  "documents": [],
  "chunks": [],
  "relation_support": [],
  "meta": {}
}

CLI Reference

Input and output

--text TEXT

Inline text to process.

Example:

rapidgraph --text "Transformer uses attention." --pretty

--input INPUT [INPUT ...]

One or more UTF-8 text files.

Examples:

rapidgraph --input input.txt
rapidgraph --input input.txt input2.txt

--output OUTPUT

Write JSON to a file instead of stdout.

Example:

rapidgraph --input input.txt --output graph.json --pretty

--pretty

Pretty-print JSON output.

Thresholds and chunking

--entity-threshold

Default: 0.35

Minimum confidence for keeping entity candidates.

--relation-threshold

Default: 0.2

Minimum confidence for keeping relation candidates.

--max-chars

Default: 600

Approximate chunk size budget.

Higher values:

• preserve more context
• may improve some relations
• increase compute cost

--chunk-mode {paragraph,sentence}

Default: paragraph

Controls chunk construction.

--chunk-overlap

Default: 1

Number of overlapping sentences preserved between neighboring chunks.

Example:

rapidgraph --input input.txt --chunk-overlap 2

Runtime mode and relation strategy

--mode {fast,balanced,quality}

Default: balanced

Controls the speed/quality tradeoff.

--max-model-spans

Default: 4

Balanced-mode only. Caps how many shortlisted spans go through REBEL.

Example:

rapidgraph --input input.txt --mode balanced --max-model-spans 6

--disable-rebel

Force heuristic-only relation extraction regardless of mode.

Example:

rapidgraph --input input.txt --mode quality --disable-rebel

Entity canonicalization scope

--entity-scope {document,corpus}

Default: document

Controls whether compatible entities can merge across files.

Use document when:

• files are independent
• names may be ambiguous across documents
• you want safer graph boundaries

Use corpus when:

• the files describe the same topic or domain
• you want one merged entity layer across the corpus
• you are building a shared graph for Neo4j or GraphRAG

Examples:

rapidgraph --input input.txt input2.txt --entity-scope document
rapidgraph --input input.txt input2.txt --entity-scope corpus

Embedding-assisted linking

These flags are optional. They are not enabled by default.

--embedding-linking

Enable embedding-assisted rescue for ambiguous entity merges and unresolved relation endpoints.

--embedding-model

Default:

sentence-transformers/all-MiniLM-L6-v2

--embedding-threshold

Default: 0.84

Cosine similarity threshold for accepting embedding-assisted merge or link candidates.

--embedding-cache-dir

Default:

.cache/extract_graph_embeddings

Embedding vectors are cached locally in SQLite form.

--embedding-max-candidates

Default: 8

Maximum number of candidates considered in embedding-assisted linking for an unresolved mention.

Example:

rapidgraph \
  --input input.txt input2.txt \
  --entity-scope corpus \
  --embedding-linking \
  --embedding-threshold 0.84 \
  --embedding-max-candidates 8 \
  --pretty

Provenance controls

--include-chunk-text

Default: enabled

Include chunk text inside the chunks array.

--no-include-chunk-text

Omit chunk text while keeping chunk metadata.

--omit-provenance-text

Alias for omitting chunk text while preserving chunk IDs and provenance structure.

Examples:

rapidgraph --input input.txt --no-include-chunk-text
rapidgraph --input input.txt --omit-provenance-text

Neo4j export

These flags are optional. Without them, the CLI only prints or writes JSON.

--neo4j-uri

Neo4j URI, for example:

neo4j://127.0.0.1:7687

--neo4j-user

Neo4j username.

--neo4j-password

Neo4j password.

--neo4j-database

Default: neo4j

Neo4j database name.

--neo4j-clean-document

Deletes matching document subgraphs before re-ingesting them.

Useful when rerunning the same files and you do not want duplicate document/chunk subgraphs.

--neo4j-embed-chunks

Generate embeddings for exported Chunk nodes and store them in Neo4j.

--neo4j-create-vector-index

Create a Neo4j vector index for Chunk embeddings. This is required for rapidgraph ask.

--neo4j-vector-index-name

Default: rapidgraph_chunk_embedding

Name of the Neo4j vector index used for chunk retrieval.

--neo4j-embedding-property

Default: embedding

Property on Chunk nodes used to store vector embeddings.

--chunk-embedding-model

Default:

sentence-transformers/all-MiniLM-L6-v2

Sentence-transformers model used for chunk embeddings.

Example:

rapidgraph \
  --input input.txt input2.txt \
  --mode quality \
  --entity-scope corpus \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --neo4j-database neo4j \
  --neo4j-clean-document \
  --neo4j-embed-chunks \
  --neo4j-create-vector-index

Logging

--log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}

Default: WARNING

Controls CLI log verbosity.

Example:

rapidgraph --input input.txt --log-level DEBUG

Recommended Flag Combinations

Fastest CPU pass

rapidgraph --input input.txt --mode fast --pretty

Best default for most users

rapidgraph --input input.txt --mode balanced --pretty

Higher-recall single document run

rapidgraph --input input.txt --mode quality --chunk-overlap 2 --pretty

Multi-file corpus merge

rapidgraph \
  --input input.txt input2.txt \
  --mode balanced \
  --entity-scope corpus \
  --pretty

Multi-file corpus with stronger ambiguous-link rescue

rapidgraph \
  --input input.txt input2.txt \
  --mode balanced \
  --entity-scope corpus \
  --embedding-linking \
  --pretty

Smaller provenance payload

rapidgraph \
  --input input.txt \
  --omit-provenance-text \
  --pretty

Neo4j ingestion with document replacement

rapidgraph \
  --input input.txt input2.txt \
  --mode quality \
  --entity-scope corpus \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --neo4j-database neo4j \
  --neo4j-clean-document \
  --neo4j-embed-chunks \
  --neo4j-create-vector-index

Neo4j Export Model

When Neo4j export is enabled, the graph currently uses:

Node labels:

• Document
• Chunk
• Entity

Relationship types:

• HAS_CHUNK
• MENTIONS
• RELATES_TO

Important detail:

• the semantic edge label such as IS_BASED_IN, USES, or DERIVED_FROM is stored as a property on RELATES_TO
• this is why Neo4j Browser may show many RELATES_TO relationships while the actual semantic relation name is visible in r.relation

Example query:

MATCH (s:Entity)-[r:RELATES_TO]->(t:Entity)
RETURN s.text, r.relation, t.text, r.evidence
ORDER BY r.relation

GraphRAG Ask Mode

rapidgraph ask lets users ask questions against a Neo4j graph that was exported with chunk embeddings.

Required setup:

pip install "rapidGraph[graphrag]"

First export data with embeddings and a vector index:

rapidgraph \
  --input input.txt \
  --mode balanced \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --neo4j-database neo4j \
  --neo4j-clean-document \
  --neo4j-embed-chunks \
  --neo4j-create-vector-index

Then ask a question:

rapidgraph ask \
  --question "What are the main relations in this document?" \
  --neo4j-uri neo4j://127.0.0.1:7687 \
  --neo4j-user neo4j \
  --neo4j-password 12345678 \
  --neo4j-database neo4j \
  --ollama-host http://127.0.0.1:11434 \
  --ollama-model llama3.2 \
  --top-k 5 \
  --graph-depth 1 \
  --max-facts 20 \
  --pretty

Ask mode retrieval flow:

• embed the question
• query the Neo4j vector index over Chunk.embedding
• expand from retrieved chunks to mentioned entities
• collect nearby RELATES_TO facts
• build a compact context packet
• ask Ollama to answer using only that context

Ask mode output contains:

• answer
• sources
• facts
• meta

Python Library Usage

Basic usage

from rapidgraph import DocumentInput, build_default_extractor

extractor = build_default_extractor(mode="balanced")

result = extractor.extract_documents(
    [
        DocumentInput(
            text="Google is based in California.",
            source="one.txt",
            title="one.txt",
        ),
        DocumentInput(
            text="Sundar Pichai leads Google.",
            source="two.txt",
            title="two.txt",
        ),
    ],
    entity_scope="corpus",
)

print(result.model_dump())

Single-document usage

from rapidgraph import build_default_extractor

extractor = build_default_extractor(mode="fast")
result = extractor.extract(
    "Transformer uses multi-head attention.",
    include_chunk_text=True,
)

print(result.model_dump_json(indent=2))

Configurable extractor construction

from rapidgraph import build_default_extractor

extractor = build_default_extractor(
    max_chars=800,
    chunk_mode="paragraph",
    chunk_overlap=2,
    mode="balanced",
    max_model_spans=6,
    embedding_linking=True,
)

Python library usage with Neo4j export

This example extracts a graph in Python and then writes it directly to Neo4j.

from rapidgraph import build_default_extractor, export_graph_to_neo4j

extractor = build_default_extractor(
    mode="balanced",
    chunk_mode="paragraph",
    chunk_overlap=1,
)

result = extractor.extract(
    """
    Google is based in California.
    Sundar Pichai leads Google.
    """,
    entity_scope="document",
    include_chunk_text=True,
)

export_graph_to_neo4j(
    result,
    uri="neo4j://127.0.0.1:7687",
    user="neo4j",
    password="12345678",
    database="neo4j",
    clean_document=True,
)

If you want to use the Neo4j helper, install the extra first:

pip install "rapidGraph[neo4j]"

Python library usage with GraphRAG ask

from rapidgraph import GraphRAGClient, Neo4jVectorRetriever, OllamaLLM

retriever = Neo4jVectorRetriever(
    uri="neo4j://127.0.0.1:7687",
    user="neo4j",
    password="12345678",
    database="neo4j",
)

llm = OllamaLLM(
    model="llama3.2",
    host="http://127.0.0.1:11434",
)

client = GraphRAGClient(retriever=retriever, llm=llm)
answer = client.ask(
    "What does the graph say about attention?",
    top_k=5,
    graph_depth=1,
    max_facts=20,
)

print(answer.model_dump_json(indent=2))

Performance and Practical Notes

CPU expectations

• fast is the cheapest mode
• balanced is usually the best practical CPU choice
• quality may be significantly slower because it runs REBEL on every chunk

First-run cost

The first run may be slower because model weights may need to be loaded or downloaded.

Hugging Face access

Some backends download models from the Hugging Face Hub if they are not already present locally.

Optional environment variables:

• HF_TOKEN
  • useful for higher rate limits
• HF_HUB_OFFLINE=1
  • useful if models are already cached locally and you want fully offline behavior

Why chunking matters

Chunking directly affects:

• relation recall
• context preservation
• runtime
• schema richness

Too-small chunks can lose relation context. Too-large chunks can increase noise and runtime. paragraph mode with a small overlap is a good default.

Troubleshooting

rapidgraph command not found

Make sure the package is installed in the active environment:

pip install rapidGraph

Slow first run

Expected if models are being downloaded or loaded for the first time.

Hugging Face warnings

Warnings about unauthenticated requests are not fatal. Set HF_TOKEN if you want authenticated Hub access.

Real PyPI vs TestPyPI

To install from TestPyPI:

pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple rapidGraph

For the public release:

pip install rapidGraph

Neo4j shows only RELATES_TO

That is expected. The semantic relationship name is stored in the relation property, not as a separate Neo4j relationship type.

Why expanded_schema can be larger than potential_schema

potential_schema is intentionally strict and compatibility-focused. expanded_schema preserves finer type detail and therefore often contains more rows.

Development

Install development dependencies:

pip install ".[dev]"

Run the main test suite:

pytest -q tests/test_extract_graph.py

Build the package:

python -m build

Validate package metadata:

python -m twine check dist/*

Publishing

TestPyPI

pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple rapidGraph

Real PyPI

pip install rapidGraph

This repository includes GitHub Actions workflows for:

• TestPyPI publishing
• real PyPI publishing

License

MIT