Document Import

SOP Studio can import existing Standard Operating Procedures from DOCX and PDF documents. An LLM extracts the decision tree structure from the document content, which you then review and refine in the visual editor.

Supported Formats

Format	Extension	Notes
Microsoft Word	.docx	Tables, lists, and headings are preserved
PDF	.pdf	Text-based PDFs; scanned documents are not supported

How It Works

1. Upload -- Drag a file onto the import area or click "Import Document".
2. Parse -- The SDK parses the document client-side using parseDocx or parsePdf.
3. Extract -- The parsed text is sent to an LLM (client-side, using your own API key) to extract a structured decision tree.
4. Review -- The extracted tree loads in the visual editor for review and adjustment.
5. Save -- Once satisfied, save the SOP to the registry.

LLM Configuration

Document extraction uses an LLM to convert unstructured text into a structured decision tree. The LLM call runs entirely client-side -- SOP Studio never sends your documents to KnowledgePulse servers.

Configure your LLM provider in the settings panel:

import type { LLMConfig } from "@knowledgepulse/sdk";

const config: LLMConfig = {
  provider: "openai",          // "openai" | "anthropic" | "ollama"
  apiKey: "sk-...",            // Your API key (stored locally)
  model: "gpt-4o",            // Model identifier
  baseUrl: undefined,          // Custom endpoint (required for Ollama)
  temperature: 0.2,           // Lower = more deterministic extraction
};

Supported Providers

Provider	Models	Local?
OpenAI	gpt-4o, gpt-4o-mini	No
Anthropic	claude-sonnet-4-20250514, claude-haiku-4-20250414	No
Ollama	Any local model	Yes

SDK Functions

The document import pipeline uses three SDK functions:

parseDocx(buffer: ArrayBuffer): Promise<ParseResult>

Parses a DOCX file and returns structured text content.

import { parseDocx } from "@knowledgepulse/sdk";

const buffer = await file.arrayBuffer();
const result = await parseDocx(buffer);

console.log(result.text);       // Plain text content
console.log(result.sections);   // Heading-based sections
console.log(result.tables);     // Extracted tables

parsePdf(buffer: ArrayBuffer): Promise<ParseResult>

Parses a PDF file and returns structured text content.

import { parsePdf } from "@knowledgepulse/sdk";

const buffer = await file.arrayBuffer();
const result = await parsePdf(buffer);

extractDecisionTree(parseResult: ParseResult, config: LLMConfig): Promise<ExtractionResult>

Sends parsed document content to the configured LLM and returns a structured decision tree.

import { extractDecisionTree } from "@knowledgepulse/sdk";

const extraction = await extractDecisionTree(result, config);

console.log(extraction.name);           // Detected SOP name
console.log(extraction.domain);         // Detected domain
console.log(extraction.decision_tree);  // ExpertSOP decision_tree array
console.log(extraction.confidence);     // 0.0 to 1.0
console.log(extraction.warnings);       // Extraction issues

Review Workflow

After extraction, the decision tree loads in the editor with visual indicators:

Indicator	Meaning
Green outline	High-confidence extraction (above 0.8)
Yellow outline	Medium confidence (0.5--0.8), review recommended
Red outline	Low confidence (below 0.5), manual editing likely needed

Review each node's properties, fix any extraction errors, and add missing connections before saving.

Tips

• Structure your documents -- SOPs with numbered steps, headings, and tables produce better extraction results.
• Use lower temperature -- A temperature of 0.1--0.2 produces more consistent decision trees.
• Review always -- LLM extraction is not perfect. Always review the result before saving to the registry.