Run the SDK without network access

knowledge works equally well on local files and on text already in memory. This tutorial covers both cases.

Before you start

• knowledge is installed.
• You have a Markdown or HTML file on disk, or a string you want to extract from.

Option A: Extract from a local file

knowledge create ./my-document.html ./my-bundle

./my-document.html is read as UTF-8. The bundle directory is created if it does not exist.

Combine with --model and --validate

knowledge --model claude-3-opus-20240229 create --validate ./my-document.md ./my-bundle

Use Python for finer control

from knowledge import Knowledge

k = Knowledge(model="gpt-4o")

# Read a local file directly
raw = Knowledge.read_source("./my-document.html")

# Extract from the in-memory text
graph = k.create("./my-document.html")

Knowledge.create accepts the file path and handles the read internally; use Knowledge.read_source only when you need the raw text for another purpose.

Option B: Extract from text already in memory

When the source is a string you have already loaded (from a database, a queue, a previous step in a pipeline), use LLMExtractor directly:

from knowledge.llm.extractor import LLMExtractor

text = """
## Installation

Run `pip install knowledge`.

## Configuration

Set `OPENAI_API_KEY` before invoking the CLI.
"""

graph = LLMExtractor(model="gpt-4o").extract(text)
for cid, concept in graph.concepts.items():
    print(f"{cid}: {concept.name}")

This skips the URL or file read entirely. The extractor splits by headings, sends each section to the LLM, and returns a KnowledgeGraph.

Option C: Write an OKF bundle from a KnowledgeGraph

Once you have a KnowledgeGraph, write it to disk with BundleSerializer:

from knowledge.kmd.bundle import BundleSerializer
from knowledge.llm.extractor import LLMExtractor
from knowledge.models import Concept, KnowledgeGraph

graph = LLMExtractor().extract("# Hello\n\nWorld.")
# ... mutate graph as needed ...
graph = graph.add_concept(Concept(id="extra", name="Extra", description="..."))

BundleSerializer().serialize(graph, "./my-bundle")

BundleSerializer does not call the LLM or perform network I/O — it is a pure filesystem operation.

When to use which

Situation	Use
Source is a URL	knowledge create <url> <dir>
Source is a local file	knowledge create <path> <dir>
Source is a string you already loaded	LLMExtractor(model=...).extract(text)
You have a graph and want to write it	BundleSerializer().serialize(graph, dir)

What you learned

• knowledge works on local files and in-memory text, not only URLs.
• LLMExtractor.extract takes a string; the SDK’s Knowledge.create adds URL and file reading.
• BundleSerializer.serialize is a pure filesystem operation with no LLM or network calls.