Introduction

When processing chapters, technical documentation, study notes, or Agent execution traces within a DSH session, a common problem is distributing the original text, conclusions, evidence, and relationships throughout the text stream. cwbcheng/dsh-knowledge-graph is a DSH Cordis plugin that splits arbitrary materials’ body text, images, or AI session execution traces into knowledge graphs. It provides bidirectional positioning between the knowledge graph and the original text, along with capabilities for evidence verification, questioning, and refinement. Below is an introduction to its capabilities, installation method, and typical usage.

What is it

dsh-knowledge-graph is maintained by cwbcheng with an MIT license.

The README states that it is a DSH dynamic Cordis plugin: one Host code runs in a Node process, and one Client code runs in the browser; overall, it is pure JS, zero dependencies, and requires no build.

package.json also declares the following dependencies and injection information:

  • peerDependencies: @deepseek-ai/cordis ^4.0.1
  • dsh.client.inject: @deepseek-ai/dsh-client-runtime
  • dsh.client.platform: web

package.json also contains scripts such as build, build:lib, build:viewer, pack:extension, kg, etc.

Prerequisites

First, confirm that the environment meets the following conditions:

  1. DSH Web is started and you enter any session:
dsh web

  1. The AI model provider is configured in the environment. The plugin defaults to the system’s current model; if not configured, it will provide a clear Chinese error message.

  2. Image splitting requires a multimodal model that supports image input. If the provider refuses the image, it will fail explicitly with model_image_unsupported.

  3. The plugin runs with the current DSH process permissions and will read the repository source code and call the model. You should verify the source code and the MIT license before installing.

Core Features

Asynchronous Text Splitting

After inputting body text such as chapters, technical documentation, or study notes, the plugin calls the LLM in background task mode, returning a knowledge graph in approximately 15–40 seconds. It supports body text up to about 1 million characters.

Direct Image to Knowledge Graph

Supports uploading 1–4 images of PNG / JPEG / WebP / GIF, with a maximum of 6 MiB per image and a total maximum of 16 MiB.

The images that can be processed include text screenshots, diagrams, flowcharts, architecture charts, statistical charts, formulas, and tables.

Image extraction first uses a multimodal model to generate a canonical visual transcription with image ranges, and then splits arrows/connections, grouping/inclusion, object correspondence, order, and legends into relationship units.

Note: Visual transcription is a model output, not pixel-level deterministic OCR. Key text, values, table cells, and connection relationships should still be verified against the original image. Current image input is for creating new knowledge graphs; incremental appending for existing graphs still uses text.

8 Node Types and 12 Relationship Types

The plugin generates 8 types of nodes:

  • fact
  • claim
  • inference
  • concept
  • definition
  • example
  • counter_example
  • rule

The plugin generates 12 types of relationships:

  • supports
  • example
  • counter_example
  • defines
  • infers
  • causes
  • is_a
  • contains
  • driven_by
  • not_is
  • analogy
  • aims_at

Bidirectional Positioning

When clicking a node in the graph, a detail card can be popped up, and it will smoothly scroll and highlight to the corresponding content unit in the original text.

When clicking a content unit in the original text, the graph will center on the corresponding node and blink it.

Graph Rendering

Graph rendering supports 4 switchable layout types:

  • Force-directed
  • Circular
  • Radial
  • Hierarchical

Graph rendering supports drag-to-pan, Ctrl+Scroll zoom, toolbar zoom, and keyboard accessibility.

Verification, Questioning, and Refinement

After generating the knowledge graph, the following operations can be performed:

  1. Quick health check: Local rule checking.
  2. AI deep review.
  3. Manual closed-loop: Adopt fix, ignore, or one-click fix for issues, and write to audit records.
  4. Active questioning: Question nodes, relationships, or the entire graph.

External Fact Checking

The following nodes can be converted to verifiable assertions:

  • fact
  • claim
  • inference
  • rule
  • definition
  • counter_example

External fact checking can use external evidence like Wikipedia, or paste domain rule sources to participate in the ruling.

Each conclusion binds an evidence link and citation; citations must be locatable in search results, and fabricated citations will be automatically downgraded.

Append Splitting and History

After having split results, you can append the next paragraph or the next piece of material. AI only extracts new content and establishes cross-paragraph relationship edges with the existing graph; the same concept does not create duplicate nodes.

Each successful split is automatically recorded:

  • Maximum of 20 entries
  • Deduplication by text
  • Single delete or clear all

The browser only saves lightweight indices such as documentId, title, count; when reviewing, it reloads the body and canonical graph from Host/SQLite.

Chapter Filtering and Candidate Review

The result area can filter graph nodes and original text paragraphs by chapter.

Candidate entities or candidate claims can be marked as:

  • Pending Review
  • Accepted
  • Rejected

Status is synced to SQLite via Host; it falls back to browser localStorage on failure.

Knowledge Graph Consumption Layer

Supports bounded structured retrieval by the following conditions:

  • Keyword
  • Node type
  • Chapter
  • grounding / entailment status

It also supports evidence-based Q&A using only the server-side canonical graph and authenticated original text.

Export

You can export the current rendered graph as HD PNG, or export:

  • Full JSON
  • Nodes CSV
  • Relations CSV

Data export is the full graph, unaffected by current chapter filtering.

Workspace and Model Selection

The floating workspace window can be dragged and resized.

The width ratio of the original text and knowledge graph, and the height of the result area can be dragged and adjusted, and the settings are remembered.

The top of the workspace and the “Trace Knowledge Graph” provide a model dropdown box where you can manually specify the model to be used for the following operations:

  • Split
  • Append
  • AI Review
  • Questioning
  • External Checking

Model selection is saved locally in the browser.

Trace Knowledge Graph

A new “Trace Knowledge Graph” tab is added to the conversation area.

It can split the current session’s user messages, tool calls, tool results, and AI replies into a knowledge graph. It supports bidirectional positioning between the graph and trace events, and also supports incremental merging for new events.

Persistent Entry

A “Knowledge Graph” button stays permanently on the right of each conversation title. The running card also has a start bar.

Installation and Enablement

First, get the source code:

git clone https://github.com/cwbcheng/dsh-knowledge-graph.git
cd dsh-knowledge-graph

Then, enter any DSH Web session and send the following message to the Agent:

请读取 dsh-knowledge-graph 仓库的 src/index.host.js 和 src/index.client.js，把这两个文件定义为 Cordis 插件的 Host 半和 Client 半，然后运行它。

The Agent will call cordis_define (define) and cordis_run (run) sequentially, and a running approval card will pop up on the interface.

After the above steps, you can open the workspace from the “Knowledge Graph” button on the right side of each conversation title or the start bar in the running card.

Typical Usage

  1. Text to Graph: Input body text such as chapters, technical documentation, or study notes in the workspace; the background will asynchronously split and generate a knowledge graph.
  2. Image to Graph: Upload 1–4 images of PNG / JPEG / WebP / GIF to generate a knowledge graph; clicking an image will locate the corresponding visual transcription paragraph.
  3. Bidirectional Positioning: Click a node in the graph to view full content, original text excerpts, and positioning buttons; click a content unit in the original text to center and blink the corresponding node in the graph.
  4. Append Splitting: After having split results, click “Append Split” to paste the next paragraph or the next material. AI extracts the new content and automatically establishes cross-paragraph relationship edges with the existing graph.
  5. Highlight Splitting: Select any text in a chat message and click the “Split into Knowledge Graph” button to automatically open the workspace and split the selected text.
  6. Verification and Refinement: Use quick health check, AI deep review, question this node, question this relationship, or external fact check the original text, and perform adopt fix, ignore, or one-click fix on issues.
  7. Retrieval and Q&A: Use the knowledge graph consumption layer to search by keyword, node type, chapter, grounding/entailment status, and perform evidence-based Q&A.
  8. Trace Graph: Open the “Trace Knowledge Graph” tab in the conversation area to decompose the complete execution trace of the current session with one click; you can also click “Append New Event” to only decompose the new part and incrementally merge it on the same revisioned document.
  9. Export: Export the current rendered graph as HD PNG, or export Full JSON, Nodes CSV, or Relations CSV.

Applicable Scenarios and Notes

Suitable for those who need to organize materials’ body text, images, and Agent session traces into retrievable knowledge graphs, and who want the graph and original text to be mutually locatable and the conclusions to be verifiable.

Note the following items:

  • The plugin runs with the current dsh process permissions and will read the repository source code and call the model; you should verify the source code and the MIT license before installing.
  • Requires a DSH Web session and an AI model provider; the plugin defaults to the system’s current model, and will provide a clear Chinese error message if not configured.
  • Image extraction requires a multimodal model that supports image input; if the provider refuses the image, it will fail explicitly with model_image_unsupported.
  • Visual transcription is a model output, not pixel-level deterministic OCR; key text, values, table cells, and connection relationships should still be verified against the original image.
  • Current image input is for creating new knowledge graphs; incremental appending for existing graphs still uses text.
  • The persistent mode saves the full text, canonical graph, and lossless checkpoint in SQLite; after refreshing, the browser only restores based on documentId/runId.
  • Only tasks left by a Host restart that have a running status are allowed to continue from the checkpoint; explicitly failed/cancelled tasks will never retry automatically.
  • Candidate entity/claim status is synced to SQLite via Host; it falls back to browser localStorage on failure.
  • In external fact checking, each conclusion binds an evidence link and citation; fabricated citations will be automatically downgraded.

Conclusion

The value of dsh-knowledge-graph lies in converting materials, images, and traces in DSH sessions into locatable, verifiable, and repairable knowledge graphs.

GitHub repository address:

https://github.com/cwbcheng/dsh-knowledge-graph

Directory entry:

https://www.skillhub.cn/plugins/cwbcheng/dsh-knowledge-graph

This directory entry comes from the plugin entry and has not been verified in the scraped material body text.