bragi-io

The Python SDK for Bragi. It turns a PDF into a bgraph — one structured, addressable graph of its sections, paragraphs, and content — and hands it back as fully typed Python objects, with the coordinates to point back at the page each piece came from.

The bgraph is the product; this SDK is one door onto it.

Install

pip install bragi-io

Python 3.9+. The only runtime dependency is httpx. On first use, the SDK fetches the bragi binary and a Java runtime automatically; later runs reuse them.

A real bgraph

import bragi

bgraph = bragi.parse_pdf("graphrag.pdf")

print(bgraph)                # <BragiGraph: 368 nodes, schema v1.2.0>
print(len(bgraph.sections))  # 54

for section in bgraph.sections:
    print(section.content.text)
    print(section.location.physical.page)

Parsing From Local to Global: A GraphRAG Approach to Query-Focused Summarization (Edge et al., Microsoft, arXiv:2404.16130, CC BY 4.0) yields 368 nodes. Every node is typed, with IDE autocomplete:

node = bgraph.sections[0]

node.content.text                    # "From Local to Global: A GraphRAG Approach to Query-Focused Summarization"
node.token_count                     # 17
node.location.semantic.path          # "1"
node.location.semantic.breadcrumbs   # ["From Local to Global: A GraphRAG Approach to Query-Focused Summarization", "From Local to Global: A GraphRAG Approach to Query-Focused Summarization"]
node.location.physical.page          # 1
node.location.physical.bounding_box  # BoundingBox(x=121.7, y=100.1, width=368.7, height=36.500008)

A semantic location (where the node sits in the tree) and a physical one (where it sits on the page) travel on every node — so you can ground an answer back on the original page.

Same code, three places the parse runs

The SDK is one interface over one pure function; where the parse runs is a one-line change. Your graph code stays the same:

# 1. Local (default) — runs the bragi binary on your machine
bgraph = bragi.parse_pdf("paper.pdf")

# 2. Self-hosted — point at your own server (see the Docker guide)
bragi.configure(url="http://localhost:8080")
bgraph = await bragi.parse_pdf_async("paper.pdf")

# 3. Hosted — point at the managed API
bragi.configure(api_key="...")
bgraph = await bragi.parse_pdf_async("paper.pdf")

Same input, same graph, whichever door you pick.

Working with the graph

bgraph.nodes                 # list[DocumentNode] — every node
bgraph.sections              # Section nodes
bgraph.paragraphs            # Paragraph nodes
bgraph.root                  # the Document root
bgraph.document_info         # metadata about the document
bgraph.structural_profile    # node counts, token + depth distributions
bgraph.schema_version        # "1.2.0"

bgraph.get_node("7962788f-d2e7-50bc-8359-c47c4b37c03d")   # by id
bgraph.nodes_by_page(1)                                    # all nodes on a page
bgraph.to_dict()                                           # the raw bgraph.json dict

Navigate the tree and render text:

parent = node.get_parent(bgraph)
children = node.get_children(bgraph)

print(section.render(bgraph, breadcrumbs=True))
# [From Local to Global: A GraphRAG Approach to Query-Focused Summarization > From Local to Global: A GraphRAG Approach to Query-Focused Summarization]
# ...the section's text, and its descendants...

Stable names

A node's id is derived from its content and its place in the tree — not from when you parsed it. Reparse after editing one paragraph and only that paragraph's id changes; every other node keeps its id. That's what lets you cache, diff, and build on a bgraph over time.

Error handling

from bragi.errors import (
    BragiError,            # base
    BragiAuthError,        # bad/missing API key (hosted)
    BragiCreditsError,     # out of credits (hosted)
    BragiProcessingError,  # the parse failed
    BragiNotFoundError,    # the bragi binary wasn't found (local)
)

try:
    bgraph = bragi.parse_pdf("document.pdf")
except BragiProcessingError as e:
    print(f"parse failed: {e}")

Documentation

• Python SDK guide — full usage, modes, rendering, error handling
• Schema reference — the bgraph.json fields

License

MIT or Apache-2.0, at your option.