Workflow: How to Build Your Own GSP Bot for AI-Assisted Research

This guide provides a step-by-step workflow to create a personalized, RAG-supported chatbot that answers Good Scientific Practice (GSP) questions about AI use in research — based on the regulations and guidelines that specifically apply to you. The bot will use GWDG ChatAI and Arcana RAG Manager.

How This Workflow Works

You will use a powerful AI assistant with internet access to work through each step. The AI does the research, asks questions and generates outputs — you validate, correct and execute. Each step provides a ready-to-use prompt to copy and paste into your AI assistant.

What You Need Before You Start

  • Access to a powerful AI assistant with internet access (e.g. Claude, ChatGPT, Perplexity, Gemini, ChatAI)
  • Access to GWDG ChatAI
  • Access to Arcana RAG Manager

Steps at a Glance

  1. Map Your Regulatory Hierarchy
  2. Identify Documents
  3. Define Contradiction and Gap Handling Rules
  4. Prepare Documents for Arcana
  5. Upload to Arcana and Configure Your ChatAI Bot
  6. Test, Validate, and Maintain Your Bot

Step 1: Map Your Regulatory Hierarchy

What This Step Does

Before you can identify which documents belong in your bot, you need to know which regulatory levels apply to you. In this step, your AI assistant will first ask you a set of questions about your role, institution, field, and funding situation. It will then use that information — together with its own internet research — to map the full regulatory hierarchy that governs your research, from the international level down to your own institution.

The output of this step is a prioritized hierarchy you can validate and correct. This hierarchy will guide every subsequent step of the workflow.

Your Prompt

Copy and paste the following prompt into your AI assistant:

I want to build a personalized RAG-supported chatbot that answers questions about Good Scientific Practice (GSP) and the use of AI in research — based on the specific regulations and guidelines that apply to me. You will help me work through a structured workflow to build this bot step by step.

We are starting with Step 1: mapping my regulatory hierarchy.

First, please ask me the following questions one by one and wait for my answers before moving on:

  1. What is your role? (e.g., doctoral researcher, postdoc, research staff, PI, professor)
  2. Which institution or institutions are relevant to you? For example, doctoral researchers may be employed by or affiliated with a research institute while simultaneously enrolled at a university — and within a university, your faculty may have additional rules that apply to you. List all institutions and, where applicable, the faculty or department.
  3. What is your academic discipline or research field?
  4. Who funds your research? (e.g., DFG, EU Horizon, BMBF, institutional funding, industry)
  5. In which country and, if applicable, which federal state or region are you based?
  6. Which of the following activities are relevant to your use of AI in research? Select all that apply: writing or publishing research papers, grant writing, data collection or processing, supervision or teaching, other (please specify).
  7. Do you publish in specific journals or for specific conferences that may have their own AI-use policies?

Once I have answered all questions, use your internet access to research the regulatory levels that apply to my situation. Then produce a numbered, prioritized hierarchy of the regulatory levels that govern my research — from the highest (e.g., EU or international) to the most specific (e.g., my institution or department). For each level, include a one-sentence description of what kind of rules it typically contains and why it is relevant for me.

End with a note on any levels where you were unable to find specific publicly available information.

Step 2: Identify Documents

What This Step Does

In this step, your AI assistant will go through each level of your validated regulatory hierarchy and identify the specific documents — guidelines, policies, laws, and codes of conduct — that are relevant to your situation and use cases. The result is a structured document list with links and access information. You will use this list in Step 4 to prepare your documents for upload.

Your Prompt

Copy and paste the following prompt into your AI assistant, then add your validated hierarchy and user profile from Step 1 at the end:

We are continuing our workflow to build a personalized GSP bot. We have completed Step 1. Below is my validated regulatory hierarchy and user profile from Step 1.

We are now working on Step 2: identifying the specific documents that are relevant to me.

For each level of my regulatory hierarchy, please use your internet access to identify the specific documents that apply to my situation. Take into account my role, my institutions, my discipline, my funders, and especially my use cases (e.g., publishing, grant writing, data processing) as described in my profile.

Before adding any document to the list, apply the following two filters in sequence. Only include a document if it passes both filters.

Filter 1 — GSP Relevance: A document is relevant only if a violation of its rules could result in at least one of the following consequences: a finding of research misconduct, rejection or retraction of a publication, loss of funding or a negative funding decision, failure or invalidation of an academic qualification (e.g., thesis rejection), or formal disciplinary proceedings at an institution. Do not include political strategy papers, consultation documents, non-binding recommendations, or general policy statements that carry no direct consequences for the researcher.

Filter 2 — Personal Relevance: Before adding any document, answer the following three questions. Include the document only if all three can be answered with yes:

  1. Does this document directly govern an activity I carry out in my research — not research in general, but specifically the activities I described in my profile?
  2. Would a violation of this document’s rules have a concrete, identifiable consequence for me personally — not for research institutions or the scientific community in general?
  3. Is this document currently in force and applicable to my situation today — not a draft, an expired version, or a document that applies only to future projects?

If a document fails either filter, do not include it. If you are unsure whether a document passes a filter, flag it with a brief note explaining your uncertainty and let me decide.

For each document that passes both filters, provide:

  1. Title and issuing organization
  2. One sentence explaining why this document is relevant to me
  3. Version date or „current as of [date]“ where identifiable
  4. Access status: openly accessible / requires login / likely behind an institutional intranet
  5. A direct link to the document, or step-by-step instructions for how to access it — including any login, VPN, library portal, or institutional access steps required
  6. A suggested file name following this convention: ##_HierarchyLevel_ShortTitle_Year.pdf — where ## is a two-digit number reflecting the document’s position in the hierarchy (01 being the highest level)

If a document cannot be publicly accessed or obtained through institutional channels, do not leave the gap empty. Instead, search for a replacement: first look for an alternative document at the same hierarchy level that covers the same topic; if none exists, identify a document from the next higher hierarchy level that addresses the same area. Explain why you are proposing the replacement.

If you are unable to find any document for a given hierarchy level, note this explicitly rather than leaving the level empty.

Structure your output by hierarchy level, in the same order as my validated hierarchy.

[Paste your validated hierarchy and user profile from Step 1 here]

What the AI Will Produce

  • A structured document list organized by regulatory hierarchy level, filtered to documents that are binding from a GSP perspective and directly relevant to your situation
  • For each document: title, issuing organization, relevance, version date, access status, link or access instructions, and suggested file name
  • Flagged documents where the AI was uncertain about filter criteria, for your decision
  • Replacement suggestions for any documents that cannot be obtained, with a brief explanation
  • Notes on hierarchy levels where no relevant document could be identified

Your Task

Review the document list. Check any documents the AI has flagged as uncertain and decide whether to include or exclude them. Remove any documents that are not relevant to your situation and add any you know are missing. Verify that links are working and that version dates are current.

For documents flagged as intranet-only or login-restricted: download these now and save them to a local folder using the suggested file names. These documents cannot be accessed by the AI in the next step.

For all other documents: save the list with links — you do not need to download these yet. The AI will access them directly in Step 4 to assess whether preparation is needed before upload.

When you are satisfied with the result, save the final document list — you will use it in Step 4.

Step 3: Define Contradiction and Gap Handling Rules

What This Step Does

Your document collection covers multiple regulatory levels that may not always agree. In this step, your AI assistant will analyze your hierarchy and documents for likely conflict zones and gaps, propose a set of default handling rules, and ask you to make decisions on edge cases. The output is a written ruleset that will be incorporated directly into your bot’s system prompt in Step 5.

Your Prompt

Copy and paste the following prompt into your AI assistant, then add your validated hierarchy and final document list from Step 2 at the end:

We are continuing our workflow to build a personalized GSP bot. We have completed Steps 1 and 2. Below is my validated regulatory hierarchy and final document list from Step 2.

We are now working on Step 3: defining the rules for how my bot should handle contradictions between documents and gaps where no document provides guidance.

Please proceed in three parts:

Part 1 — Conflict Analysis: Analyze my hierarchy and document list for likely conflict zones — situations where rules from different levels may contradict each other. List the most probable conflict scenarios relevant to my use cases.

Part 2 — Default Rules: For each conflict scenario, propose a default handling rule. Additionally, propose the following general default rules and ask me to confirm or modify each one:

  • When two rules conflict, the rule from the higher hierarchy level takes precedence.
  • When a more specific rule exists at a lower level (e.g., a faculty policy that is stricter than the university policy), the more specific rule applies.
  • When no document in the collection addresses a situation, the bot will say so explicitly and refer the user to the most relevant contact or body for clarification.

Part 3 — Edge Cases: Identify situations where the default rules are insufficient and a human decision is needed. For each edge case, ask me one question at a time and wait for my answer before moving on.

Once all rules and edge case decisions are confirmed, produce a final ruleset formatted as a numbered list of „If [situation], then [action]“ rules — written in a way that can be copied directly into a system prompt. Include a standard disclaimer rule at the end: a formulation the bot will use whenever it is uncertain about the currency or completeness of its information, directing the user to verify with their institution.

[Paste your validated hierarchy and final document list from Step 2 here]

What the AI Will Produce

  • An analysis of likely conflict zones based on your hierarchy and documents
  • A set of proposed default rules for you to confirm or modify
  • Targeted questions on edge cases requiring your decision
  • A final numbered ruleset in „If [situation], then [action]“ format, ready to be copied into a system prompt
  • A standard disclaimer rule the bot will use when uncertain

Your Task

Review and confirm or adjust the proposed default rules. Answer the AI’s edge case questions one by one. When the final ruleset is produced, review it carefully — these rules will directly govern how your bot behaves. Save the final ruleset, as you will need it in Step 5.

Step 4: Prepare Documents for Arcana

What This Step Does

Before uploading your documents to Arcana, they need to be in a format that Arcana can process well. Arcana accepts PDF, DOCX, TXT, and Markdown files and converts them automatically into structured Markdown using its Docling engine. However, the quality of that conversion depends on the quality of the source file. In this step, your AI assistant will access each document via its link, assess its quality, and tell you whether preparation is needed and, if so, exactly what to do.

What Arcana Needs

  • Supported formats: PDF (with embedded text), DOCX, TXT, Markdown
  • Recommended file size: under 50 MB per file (maximum: 300 MB)
  • Best results: well-structured files with clear headings, short paragraphs, and no complex multi-column layouts
  • Avoid: scanned PDFs without OCR, password-protected files, heavily formatted files with tables spanning multiple pages

Your Prompt

Copy and paste the following prompt into your AI assistant, then add your final document list from Step 2 at the end:

We are continuing our workflow to build a personalized GSP bot. We have completed Steps 1, 2, and 3. Below is my final document list from Step 2.

We are now working on Step 4: preparing my documents for upload to Arcana, a RAG system that accepts PDF, DOCX, TXT, and Markdown files and converts them automatically into structured Markdown during indexing.

For each document in my list that has a public link, please access the document and assess whether preparation is needed before upload. For each document, provide:

  1. Format assessment: is the file in a supported format (PDF with embedded text, DOCX, TXT, or Markdown)?
  2. Quality assessment: is the document well-structured with clear headings and short paragraphs, or does it have complex layouts, multi-column formatting, or other features that may reduce retrieval quality?
  3. Size assessment: is the file likely under 50 MB?
  4. Preparation recommendation: one of the following:
    • Ready to upload — no preparation needed; provide the direct download link
    • OCR needed — the document appears to be a scanned PDF without embedded text; recommend a specific free tool to apply OCR (e.g., OCRmyPDF, ILovePDF, Adobe Acrobat Online)
    • Convert to Markdown — the document has a complex layout that is likely to convert poorly; provide a clean Markdown version of the document’s content if you have full access to it via the link, or explain why conversion is recommended and how the user can do it manually
    • Other preparation needed — describe what needs to be done and why

If you cannot access a document via its link, flag it explicitly. In that case, ask me to download the document manually and describe its format and structure so you can still provide a preparation recommendation.

For documents I have already downloaded locally (intranet or login-restricted), I will describe their format and structure to you directly — please provide a preparation recommendation based on my description.

Structure your output by document, using the file names from Step 2.

[Paste your final document list from Step 2 here]

What the AI Will Produce

  • A per-document assessment of format, quality, and file size
  • A clear preparation recommendation for each document: ready to upload, OCR needed, convert to Markdown, or other
  • Direct download links for documents that are ready to upload as-is
  • Where applicable: tool recommendations for OCR, or a ready-to-use Markdown version of the document
  • Explicit flags for any documents the AI could not access, with a request for manual download and description

Your Task

Work through the AI’s recommendations document by document. For documents flagged as ready to upload, download them now using the provided link. For documents requiring OCR or conversion, follow the instructions provided. If the AI could not access a document, download it manually, describe its format and structure to the AI, and follow its recommendation.

When all documents are prepared and saved in your local folder with the file names from Step 2, you are ready to move on to Step 5.

Step 5: Upload to Arcana and Configure Your ChatAI Bot

What This Step Does

In this step you will upload your prepared documents to Arcana, index them, generate a complete system prompt for your bot, and configure it in ChatAI. By the end of this step your personalized GSP bot will be ready to use.

Part A: Upload Your Documents to Arcana

Follow these steps in the Arcana Manager:

  1. Log in with your Academic Cloud or university SSO account.
  2. Go to the Arcanas tab and click + Add New Arcana.
  3. Enter a clear, descriptive name for your Arcana (e.g., GSP-AI-Bot-YourName).
  4. Optional: enable Secure Arcana Name if you want to restrict access to people you share the link with explicitly.
  5. Click Upload Files and select all the documents from your local folder. You can select multiple files at once.
  6. Once all files are uploaded, click Index Generation and confirm. Wait until the status of all files changes to Indexed (shown in green).
  7. Click Access Link to generate your Arcana’s access link and note down the Arcana ID. Save both — you will need them to share your bot and to configure ChatAI.

For detailed guidance, refer to the official Arcana getting started documentation.

Part B: Generate Your System Prompt

Copy and paste the following prompt into your AI assistant, then add your user profile from Step 1 and your validated ruleset from Step 3 at the end:

We are continuing our workflow to build a personalized GSP bot. We have completed Steps 1 through 4. Below are my user profile from Step 1 and my validated ruleset from Step 3.

We are now working on Step 5: generating the system prompt for my bot.

Please generate a complete system prompt for a RAG-supported chatbot that answers questions about Good Scientific Practice and the use of AI in research. The system prompt should:

  1. Define the bot’s role and purpose clearly, tailored to my specific profile (role, institutions, discipline, funders, and use cases)
  2. Incorporate all rules from my validated ruleset from Step 3 exactly as formulated, including the disclaimer rule
  3. Instruct the bot to always respond in the language of the user’s question — if the question is in German, respond in German; if in English, respond in English; and so on for other languages
  4. Define a helpful, professional, and precise tone — the bot should be direct and concise, avoid unnecessary hedging, and always cite the source document and hierarchy level it is drawing from
  5. Instruct the bot to structure its answers clearly: first the direct answer, then the source, then any caveats or recommendations to verify with the institution
  6. Instruct the bot to explicitly say when a question falls outside the scope of its documents, rather than guessing

Output the system prompt as a single, clean block of text that I can copy and paste directly into ChatAI.

[Paste your user profile from Step 1 here]

[Paste your validated ruleset from Step 3 here]

Part C: Configure Your Bot in ChatAI

Follow these steps in GWDG ChatAI:

  1. Open ChatAI and go to the advanced settings.
  2. Enter your Arcana ID in the Arcana field.
  3. Select the recommended model: Qwen 3 30B A3B Instruct.
  4. Set temp to 0.0 and top_p to 0.05 for optimal retrieval accuracy.
  5. Paste your generated system prompt into the System Prompt field.
  6. Send a test question to confirm the bot is responding correctly and drawing from your documents.

What the AI Will Produce

  • A complete, ready-to-use system prompt tailored to your profile, use cases, and validated ruleset

Your Task

Upload and index all documents in Arcana. Generate your system prompt and review it carefully before pasting it into ChatAI. Configure ChatAI with your Arcana ID, the recommended model, and the correct parameter settings. Run a test question and check that the bot responds correctly, cites its sources, and applies your ruleset. Save your Access Link — you will need it in Step 6.

Step 6: Test, Validate, and Maintain Your Bot

What This Step Does

In this step you will systematically test your bot with a set of targeted questions, evaluate the quality of its responses, and document any issues. Your AI assistant will generate the test questions for you based on your profile, use cases, and ruleset. At the end of this step your bot is either ready to use or you have a clear list of issues to address.

Part A: Generate Test Questions

Copy and paste the following prompt into your AI assistant, then add your user profile from Step 1 and your validated ruleset from Step 3 at the end:

We are continuing our workflow to build a personalized GSP bot. We have completed Steps 1 through 5. Below are my user profile from Step 1 and my validated ruleset from Step 3.

We are now working on Step 6: testing and validating my bot.

Please generate a set of test questions for my bot. The questions should cover the following categories:

  1. Standard use case questions — typical questions a researcher in my situation would ask about GSP and AI use, covering each of my use cases (publishing, grant writing, data processing, etc.)
  2. Conflict and edge case questions — questions that are designed to trigger the contradiction and gap handling rules from my ruleset
  3. Out-of-scope questions — questions that fall clearly outside the scope of my documents, to test whether the bot correctly says it does not know
  4. Disclaimer trigger questions — questions where the bot should apply the disclaimer rule

For each question, provide the expected answer in one or two sentences — what a correct, well-behaved bot response should contain. Format the output as a table with three columns: Question, Expected Answer, and Result (leave the Result column empty — I will fill it in during testing).

[Paste your user profile from Step 1 here]

[Paste your validated ruleset from Step 3 here]

Part B: Run the Tests

Open your bot via the Access Link from Step 5 and ask each question from the test table. For each response, fill in the Result column with one of the following:

  • OK — the response matches the expected answer, cites the correct source, and applies the ruleset correctly
  • Check — the response is partially correct or unclear; needs closer review
  • Fix — the response is incorrect, missing a source citation, or fails to apply the ruleset

For any result marked Fix, note what went wrong. Common issues and how to address them:

  • Bot gives wrong or incomplete answer → check whether the relevant document is correctly indexed in Arcana; re-upload if needed
  • Bot does not apply a rule correctly → revise the system prompt and re-paste it into ChatAI
  • Bot does not say it does not know when it should → add a more explicit instruction to the system prompt

Keeping Your Bot Up to Date

Your bot’s answers are only as current as the documents in your Arcana. Whenever a document in your collection is updated — for example, when your institution revises its AI policy, your funder publishes new guidelines, or a journal updates its AI-use rules — you will need to replace the outdated file in Arcana with the new version and re-run the index generation.

Your Task

Run all test questions, fill in the Result column, and address any issues marked Fix. Once all results are OK or Check and you are satisfied with the bot’s performance, your bot is ready to use. Save your completed test table for future reference.