DORA Admin Manual

🛠️ Add a New Tool in DORA from Admin page

To add a new tool in DORA and enable its use in template-based document generation, you need to update several components in the admin panel. Follow these steps:

1. Define the Tool

Navigate to: /admin/kernel/tool/

In this section:

• Click the Add tool button to create a new tool entry. Ensure the tool name entry exactly matches the implementation name used in the DORA codebase.
• Provide a clear description of the tool for LLMs — this helps the language model understand the tool's purpose and how it can assist in generation.
• Set the Source to "local" to indicate that the tool is handled in the local DORA environment (instead of with MCP).
• Define the argument schema required by the tool, including all necessary input fields.
• Skip the Path and Method fields — these are only required for tools executed via the MCP server.

2. Add Tool Instructions

Navigate to: /admin/kernel/toolpromptinstruction/

This step enables dynamic insertion of tool instructions into prompt templates.

• Click the Add Tool Prompt Instruction button to create a new instruction entry for the tool.
• These instructions will be automatically injected into the template prompt when:
• The tool is specified in a section within the template.
• The user chooses the template and selects the tool in the interface to generate a document.

3. Add the Tool to a Tool Group (Prompt Design and Selection)

Navigate to: /admin/kernel/toolpromptinstructiongroup/

DORA dynamically integrates tools into prompt templates using placeholder-based substitution. In this section:

• Add the tool to an existing tool group (e.g., scientific_search_tools) or create a new group by clicking the "Add tool prompt instruction group" button.
• Add an "empty state" message, which will be displayed in the prompt when no tools from the group are selected.

Add the name placeholder: {activated_tools|group_name|names}

• → Inserts a list of active tool names into the prompt.

Add the instruction placeholder: {activated_tools|group_name|instructions}

• → Inserts the combined instruction texts from all active tools in the group.

• These placeholders are dynamically replaced during generation based on the selected tools.

• This section also controls tool visibility and toggling in the user interface, also known as the Resources and Team of Agents.

Prompt Placeholder Example

Before substitution:

Use {activated_tools|scientific_search_tools|names} when needed.

{activated_tools|scientific_search_tools|instructions}

After substitution (with tools enabled):

Use pubmed_abstract_similarity_search_tool and web_search_tool when needed.

Instructions for pubmed_abstract_similarity_search_tool: 1. Questions to the pubmed_abstract_similarity_search_tool should...

Instructions for web_search_tool: 1. Questions to the web_search_tool should be short and...

Validation Rules

• {...|names} and {...|instructions} must always appear together.

• All tools listed must:

• Exist in /admin/kernel/tool/
• Have a valid instruction in /admin/kernel/toolpromptinstruction/
• If a prompt uses placeholders, at least one tool in the group must be selected for successful substitution.

4. Configure Tool Group Display and Visibility in UI

Navigate to: /admin/kernel/templateagent/

This configuration controls how tools appear in the interface. To display a tool or group of tools in the UI:

• Add an emoji for visual grouping (optional but recommended).

• Set a display name and provide a clear description.

• Attach the relevant tool(s) to the agent.

• This allows users to toggle tools on/off directly from the interface when configuring generation.

5. Final Step: Add Tool to Template Configuration

To enable tool execution during generation, the new tool must be added to the corresponding section(s) of the template configuration, which is described in the DORA Configuration File.

See also: DORA Configuration File – Section Object → tools key

1. Open the target template in the Admin panel.

2. Locate the tools key inside the desired section object.

3. Add the tool name to the list of tools for that section.

You can assign the tool to:

• A specific section of the template, or

• All sections, depending on its scope and relevance.

🧩 Example Template Section Snippet (from DORA Configuration File)

{
    "section_name": "Background",
    "tools": [
        "web_search_tool",
        "your_new_tool_name"
    ]
}

Each section's configuration is stored in the JSON structure described in the DORA Configuration File, where section_name, tools, input, and other properties define how the model operates on that section.

Once this step is complete, your new tool will be fully integrated into the DORA generation system and ready for use by end users.

🧪 Load custom Template

To load a custom template into the system, follow these steps:

1. Access the Admin Panel

• Go to /admin

• Log in using the superuser credentials you created earlier.

2. Load or Create a Template

You have two options:

Option A: Use Predefined Templates and Modify the Config

• In the Admin panel, navigate to Templates under the Kernel section.

• Check if any default templates are available.

• Select a template and adjust the configuration as needed.

Option B: Create Your Own Template

1. Navigate to /admin/kernel/template/ and click "Add Template".

1. Upload your .json configuration file.

💡 Ensure the JSON structure is correct. You can validate it using https://jsoneditoronline.org.

1. Click "Save" and wait for the system to validate your template.

2. Once validated, make the template available for use:

   • Return to the template list at /admin/kernel/template/.

   • Select your template.

   • Enable the "Is online" checkbox.

3. (Optional) Set the display order of your template on the selection page using the "Order" field.

DORA prompts for LLM features

🟢 Configs for Document generation

1. dynamic_section_prompts

Purpose: Enables dynamic section creation and detailed prompt generation from a user-defined Research Plan. When enabled ("dynamic_section_prompts": true in the template configuration), DORA intelligently splits a research plan into meaningful sections, each with its own title, logic, and generation instructions.

How it works:

• The generate_sections_template guides the system to analyze the Research Plan and divide it into major thematic sections.

• For each section, the LLM system generates:

• slug: a short, URL-safe identifier

• title: human-readable section name

• depends_on: list of section slugs this section depends on

• tools: list of tools required to generate this section

• prompt: detailed, step-by-step instructions for generation

• expected_output_instructions: guidelines on how the output should be formatted (e.g., length, structure)

• The separate_custom_data_template is used to distribute optional CUSTOM TEXT across the generated sections. This ensures all external data is correctly matched to the most appropriate section by content similarity.

When to use: When the document structure is unknown in advance and should be derived from the research plan.

2. word_filtering_feature

Purpose: Performs post-processing of generated sections to eliminate predefined blacklisted phrases and improve human-like quality and natural language fluency, especially for final polishing before document delivery.

How it works:

• The prompt instructs the model to act as a university professor refining a draft.

• It preserves structure, citations, and formatting while replacing or removing phrases such as:

• "In conclusion", "Moreover", "Extremely", etc.

• If a section contains fewer than 3 sentences, the system will leave it unchanged.

• The refined version must maintain:

• All original citations

• Section formatting (e.g., lists, bold)

• Coherent and professional tone

• No blacklisted phrases

When to use: As a quality control step after section generation to align the tone and language with academic writing standards.

3. polish_document_prompts

Purpose: Refines the entire document structure and wording in JSON format to improve flow, coherence, and readiness for academic publication.

How it works:

• Operates on the results key of each section/subsection within a JSON structure.

• Enhances:

• Sentence flow
• Section transitions
• Scientific clarity

• Language sophistication

• Strict rules enforced:

• No repetition of facts or sentences
• Avoids banned phrases (e.g., "In summary", "Delve", "Overall")
• Explains abbreviations only once
• Maintains all citation formats and placements
• Avoids starting multiple paragraphs with the same word

When to use: In the final step of document generation to ensure text is polished, logically cohesive, and meets publication standards.

4. mermaid_diagram_prompts

Purpose: Generates or edits a Mermaid.js diagram to visually summarize the document content, typically in the form of a flowchart or graphical abstract.

How it works: This configuration includes several prompts for different use cases:

• summary_prompt: Extracts key relationships, mechanisms, and workflows from the full document and creates a structured summary for flowchart generation.
• assistant_prompt: Converts the summary into Mermaid.js code, applying visual and semantic formatting guidelines (e.g., pastel colors, scientific layout).
• diagram_reviewer_prompt: Validates and corrects Mermaid.js syntax.
• summary_prompt_detailed: Offers more structured and section-based summarization for advanced diagrams.
• Other prompts (e.g., system_prompt_for_mermaid_*) are used to generate diagrams of different types (flowchart, timeline, state diagram, etc.) with design constraints and hierarchy.

When to use: To produce graphical abstracts or visual summaries to accompany scientific documents, especially useful for publication or presentation.

🧩 Other LLM-Related Features

This section provides a detailed overview of the configuration schemas used to support additional LLM-based features in the DORA platform, including:

1. polish_prompts – for text polishing after user edits.
2. ai_actions_prompts – for editing highlighted text (shorten, extend, or custom actions).
3. ai_review_prompts – for automated quality evaluation of generated scientific documents.

1. polish_prompts

Purpose:

The polish_prompts configuration defines how the LLM should rewrite a section of a scientific document after a user has manually edited it, focusing on improving grammar, coherence, style, and alignment with prior sections of the document.

Usage Scenario:

When the user manually changes a section and wants to "Polish" the updated text using LLM assistance.

Configuration Schema:

{
    "human_message": "Write the {title} section using the Initial draft.\\n### Initial draft\\n{customized_section_results}\\n\\nMain Instructions:\\n1. Determine if there are Prior sections ... \[truncated for brevity\]",
    "system_message": "You are a scientific writer. You are given the draft of the initial section. Your task: ... \[truncated for brevity\]"
}

Key Highlights:

• Conditional Logic: Behavior changes based on the presence/absence of prior sections.
• Strict Citation Retention: Must retain citations if present.
• Format Preservation: Maintains original structure (e.g., plain text or subsections).
• Coherence Enforcement: Uses prior sections as the "source of truth" to adjust the draft.

2. ai_actions_prompts

Purpose:

Handles user-invoked edits on highlighted text segments within the document. Supported actions include:

• make_longer: Extend the text with more details and citations.
• make_shorter: Condense the text to its essential points.
• custom_prompt: Apply a custom action described by the user (e.g., simplify, rephrase, clarify).

General Schema:

Each action has:

• A human_message: Instructions shown to the LLM on how to perform the transformation.
• A system_message: Defines the LLM's identity and role.

make_longer

Extends a passage to 2–3 times its original length with additional detail, smooth flow, and citations from metadata.

Key Constraints:

• Add no more than one paragraph (preferably inline expansion).

• Focus only on core topics from the initial text and metadata.

Insert citations using the format:

(BIB_ID:\[BIB_ID\], CHUNK_ID:\[CHUNK_ID\])

make_shorter

Condenses the original text, aiming to reduce length by approximately 3 times while preserving meaning and all citations.

Rules:

• Do not remove any valid citations.
• Only shorten if the initial text is sufficiently long.

custom_prompt

Performs a user-defined action (e.g., "simplify", "make more formal").

Dynamic Parameters:

• {custom_action}: Action type defined at runtime.
• Adjusts output while respecting citation and structure guidelines.

3. ai_review_prompts

Purpose:

Runs automated document review with scoring and suggestions using the LLM. Provides structured feedback across multiple metrics tailored to scientific writing.

Components:

1. general_evaluation

2. Overall Impression: Holistic assessment of the text's strengths, weaknesses, and completeness.

3. detailed_evaluation_metrics

Includes four granular scoring categories:

• language_and_style – grammar, tone, sentence structure.
• content_and_relevance – accuracy, depth, and usefulness of content.
• readability_and_structure – clarity, flow, and document structure.
• argumentation_and_evidence – strength and logic of claims, source usage.

Each has:

• A name and description.
• Instructional prompt for evaluation.
• Format of output in JSON with score, score_explanation, strengths, and numbered suggestions.

Example Output Format:

{
    "score": 7,
    "score_explanation": "The text is generally well-structured but has occasional grammar issues.",
    "strengths": "Concise writing with a clear logical structure in the Results section.",
    "suggestions": {
        "1": {
            "text": "Improve clarity in the Introduction by simplifying the definition of 'oncogenic pathways'.",
            "seriousness": "medium"
        },
        "2": {
            "text": "Fix inconsistent citation formatting in the Discussion section.",
            "seriousness": "low"
        }
    }
}

Integration Notes

• These configuration objects are injected into the LLM orchestration backend dynamically based on the button/tool activated by the user.
• expected_output and text_context are dynamically inserted by the backend pipeline during prompt execution.
• Citation formatting rules are consistent across all prompt types and enforced via template logic.