A Source-Grounded Conversational Research Assistant for Avicenna's Canon of Medicine

Conversational AI · Historical Knowledge Systems · Prompt / Workflow Engineering · Digital Humanities · Persian-Language Education

Year: 2026
Original delivery: Public Custom GPT in ChatGPT
Primary corpus: Five volumes of the Canon of Medicine, supplied as a consolidated knowledge file (master.txt)
Role: Independent product concept, knowledge-base preparation, conversational workflow design, instructions, scope definition, configuration, testing and public release
Domain: History of medicine, scholarly interpretation and non-clinical education

An interactive, Persian-language educational encyclopedia that uses a supplied corpus of Avicenna's Canon of Medicine to examine fictional case studies through the terminology and explanatory frameworks of classical medicine. It does not diagnose, treat, prescribe, or present historical theories as modern medical conclusions.

1. Project Overview

Classical Medicine — Educational Encyclopedia is a publicly released Custom GPT designed to make a large classical medical corpus accessible through structured conversation rather than a conventional chapter-by-chapter reading experience.

The foundation is the five-volume Canon of Medicine attributed to Avicenna, consolidated by the creator into a single knowledge resource named master.txt. The product combines this reference material with a purpose-built instruction set governing how hypothetical historical case studies are gathered, confirmed, discussed and cited.

The user can provide a fictional situation, and the assistant organizes its details before considering how related topics appear in the classical source. Its answers distinguish direct references, historical concepts, interpretive parallels and unsupported exact correspondences.

The engineering challenge was not simply uploading a book to ChatGPT. It was defining a useful, repeatable interaction pattern while controlling scope, source attribution, uncertainty and the tendency to turn medical vocabulary into contemporary health guidance.

2. The Problem

Large historical works pose several accessibility problems for readers and researchers:

• Relevant material can be distributed across different books, sections and chapters.
• Readers often begin with a question or scenario rather than a known page or title.
• The vocabulary of classical medicine belongs to a historical theoretical system, not contemporary clinical medicine.
• A generative assistant can confuse a plausible interpretation with a documented passage if it is not instructed to separate them.
• A single, overly confident answer is less valuable than a traceable explanation whose limitations remain visible.

The project responds to these problems with a bounded conversational reference experience.

3. Product Objective

The intended user experience is to:

1. Accept a hypothetical situation for academic or historical study.
2. Request limited contextual data relevant to describing the scenario.
3. Summarize the supplied information and seek confirmation.
4. Decompose the topic into concepts that can be investigated in the classical source.
5. Refer to relevant books, sections, chapters or internal structural identifiers where available.
6. Explain relationships through historical terminology without converting them into a diagnosis.
7. State when the requested exact correspondence is absent from, or cannot be substantiated by, the supplied text.

The project deliberately excludes diagnosis, treatment, prevention plans and personalized medical guidance.

4. Knowledge Foundation

Source corpus

• Historical work: Avicenna's Canon of Medicine (Al-Qanun fi al-Tibb).
• Scope supplied by the creator: the five-volume work.
• Configured knowledge asset visible in the Custom GPT: master.txt.
• Language of the conversational product: Persian.

The product is a knowledge-configured Custom GPT, not a separately trained foundation model. Its creator defined its purpose, uploaded the reference material, supplied interaction instructions and tested the resulting behavior.

The exact completeness, edition, textual provenance and chapter segmentation of master.txt should be documented separately before making stronger claims about source coverage.

Source-handling rules

The instruction design requires the assistant to:

• Prefer the uploaded textual material when responding.
• Refer to the source by book, section or structural identifier where possible.
• Explicitly disclose when an exact statement cannot be found.
• Avoid fabricated quotations, page numbers and precise attributions.
• Use general explanations of the historical framework only when they are clearly identified as such.

These are configured behavioral requirements; sustained compliance across the complete corpus requires a broader evaluation than a single example.

5. Conversational Architecture

flowchart TD
    A[Hypothetical academic case study] --> B[Limited contextual intake]
    B --> C[Structured case summary]
    C --> D{User confirms summary?}
    D -->|Correction needed| B
    D -->|Confirmed| E[Decompose themes]
    E --> F[Consult configured historical corpus]
    F --> G[Identify relevant source sections]
    G --> H{Exact match supported?}
    H -->|Yes| I[Explain documented historical concepts]
    H -->|Not established| J[Disclose evidence gap]
    I --> K[Non-clinical scholarly synthesis]
    J --> K

Stage A — Case intake

The user describes a fictional person or situation. Instructions permit limited context such as age, sex, main complaints and overall timing/progression. The assistant must not solicit details as if conducting a real clinical examination.

Stage B — Explicit confirmation

Before interpretation, the assistant produces a structured summary and asks whether it accurately represents the scenario. This is a human-in-the-loop checkpoint designed to reduce errors introduced by ambiguous user input.

Stage C — Historical topic decomposition

Once confirmed, the scenario is separated into themes that can be discussed against the reference corpus: for example, concepts of bodily faculties, sleep and wakefulness, head-related terminology, appetite and temporal progression.

Stage D — Source-referenced interpretation

Rather than forcing several observations into one named historical condition, the assistant can discuss multiple relevant source locations and explain the interpretive relationship between them.

Stage E — Bounded synthesis

The response retains a historical perspective and makes clear that this framework does not describe current medical understanding or establish a diagnosis.

6. Interaction Design and Safety Boundaries

The product's instruction set explicitly defines its role as an educational and historical encyclopedia.

This makes the system an exploration interface for historical ideas, not a medical assistant.

7. Practical Evaluation — Observed Case Study

A complete interaction was provided for portfolio assessment. It demonstrates the product's workflow on one fictional scenario.

Submitted scenario

A fictional 38-year-old man described:

• approximately six weeks of tiredness toward the end of the day;
• mild heaviness in the head;
• light sleep;
• reduced appetite;
• gradual onset and relative stability in the preceding two weeks.

Observed Step 1 — Structured summary

The Custom GPT restated the scenario in a table with the person's sex, age, duration, onset, each stated complaint and progression. It then requested confirmation before continuing.

Observed Step 2 — Confirmation

The user replied affirmatively. The assistant proceeded to historical analysis without inventing additional personal details.

Observed Step 3 — Thematic analysis

The answer separated the study into related historical topics:

1. Tiredness and the classical concept of bodily faculties.
2. Sleep, wakefulness and their place in the historical explanatory system.
3. The term thiqal al-ra's / head heaviness and related discussions.
4. Diminished appetite within discussions of natural faculties.
5. Gradual onset and temporal descriptions of changing conditions.

Observed Step 4 — Source identifiers

The output included structural references such as:

Evidence limitation: These are identifiers visible in the generated response. Their exact correspondence to the underlying edition and text in master.txt has not yet been independently verified. They should be described as produced citations rather than audited citations until checked.

Observed Step 5 — Evidence-gap disclosure

The assistant expressly indicated that relevant separate passages existed but that it had not found one passage describing the exact four-part symptom combination as a unified historical pattern.

This is a particularly useful observed behavior: the system did not need to invent a neat historical label to complete the exercise.

Observed Step 6 — Non-clinical conclusion

It concluded with a conceptual comparison of the separate historical themes and avoided a definitive diagnosis, an individualized treatment plan or modern medical equivalence.

8. What the Demonstration Actually Establishes

This one test supports a limited but meaningful set of observations:

It does not yet establish: corpus-wide factual accuracy, perfect retrieval, absence of hallucinations, consistency across many sessions, medical reliability or source-faithful quotation at scale.

This distinction makes the portfolio claim credible and repeatable.

9. Original Engineering Contribution

My contribution consisted of:

• defining the independent educational-product concept;
• preparing and supplying the historical knowledge material;
• setting up and publishing the Custom GPT;
• designing a staged intake / confirmation / explanation workflow;
• writing instructions for source priority and historical interpretation;
• defining failure behavior when a direct reference cannot be substantiated;
• imposing explicit non-clinical boundaries;
• selecting Persian conversation starters;
• testing the product against hypothetical scenarios;
• preparing a path for migration to a reusable plugin.

The project uses ChatGPT's existing model capability. It does not claim the underlying LLM was trained, fine-tuned or developed from scratch by me.

10. Technical Architecture

Historical Source Corpus
   Five-volume Canon of Medicine
               │
               ▼
      Consolidated master.txt
               │
               ▼
         Custom GPT Setup
         ├── Role definition
         ├── Workflow instructions
         ├── Knowledge priority
         ├── Attribution rules
         ├── Uncertainty handling
         ├── Non-clinical scope
         └── Persian starters
               │
               ▼
          User Interaction
         ├── Case intake
         ├── Summary confirmation
         ├── Topic interpretation
         └── Source-oriented synthesis
               │
               ▼
     Historical Educational Output

Engineering pattern

The product is best characterized as a knowledge-grounded, instruction-configured conversational application. Its differentiator is not model-training infrastructure but controlled interaction design around a specialist source corpus.

11. Research Value

Potential use cases include:

• guided orientation within a large historical medical work;
• studying how different concepts recur across source sections;
• understanding relationships among classical terms;
• comparing the structure of historical explanations;
• generating a source-navigation starting point for further manual reading;
• teaching the history of medicine in Persian.

The source text must remain the authority for scholarly claims. A model-generated explanation is a research aid, not independent primary-source evidence.

12. Current Limitations and Validation Roadmap

A. Citation verification

Create a small gold-standard test set from master.txt. For each question, verify that returned chapter IDs, titles and paraphrases match the actual supplied edition.

B. Evidence traceability

Where short excerpts can be used appropriately, show:

Claim / historical topic
       ↓
Book + section + structural identifier
       ↓
Brief original-source excerpt
       ↓
Persian explanation
       ↓
Explicit distinction: citation vs interpretation

C. Corpus preparation

Document the source edition, provenance, chapter hierarchy, encoding, segmentation, any editorial transformations and completeness of the consolidated master.txt file.

D. Regression tests

At minimum, evaluate:

• an ordinary hypothetical scenario;
• an ambiguous multi-theme scenario;
• an intentionally unsupported quotation request;
• a request for a fabricated page number;
• a prompt attempting to extract personalized treatment advice;
• a scenario requiring the model to say that the source contains no exact match;
• a follow-up correction to the initial case summary.

E. Product usability

Keep the confirm-before-analysis workflow, but consider an optional scholarly mode that allows users to explore one named chapter directly, without inventing a fictional case first.

13. Migration / Product Continuity

As of September 2026, the original publicly accessible Custom GPT displays a migration notice. The intended successor is a Plugin combining reusable instructions and reference material.

Important operational distinction: Migration produces a replacement plugin; it does not automatically preserve the original GPT's public-sharing configuration. A migrated personal plugin starts private, and wider publication requires a separate process.

Before migration, preserve:

1. The original master.txt as an independent local file.
2. The complete Instructions text.
3. Name, description, icon and conversation starters.
4. Screenshots of the current configuration and public page.
5. At least one complete baseline conversation, including the confirmed case-study example above.
6. The latest published version of the GPT; unpublished drafts are not used by migration.

After migration, compare the replacement plugin with the saved baseline and verify that its instructions and knowledge files have been retained. The source text, prompt design and publication documentation should remain under independent project control rather than existing only inside one product interface.

Official migration reference: OpenAI — Custom GPT Retirement and Migration FAQ.

14. Suggested Website Hero

Classical Medicine — Educational Encyclopedia

Avicenna's Canon of Medicine · Conversational AI · Digital Humanities

A publicly released Persian-language Custom GPT that transforms a five-volume historical medical reference into a structured, source-oriented educational conversation. The assistant gathers fictional scenarios, verifies their details, decomposes them into historical themes and produces cautious explanations using classical terminology and document references, while explicitly excluding diagnosis, prescriptions and personalized medical guidance.

15. Website Project Description

Classical Medicine — Educational Encyclopedia began as an experiment in making Avicenna's Canon of Medicine accessible through interactive research rather than static reading. I consolidated the reference corpus into a knowledge file and developed the Custom GPT's behavior around a structured academic workflow: scenario intake, confirmation of the user's facts, historical topic decomposition, source-oriented explanation and uncertainty disclosure.

In a documented demonstration, the system analyzed a fictional multi-symptom scenario by connecting its themes to separate parts of the classical text, producing internal book/section identifiers and explicitly declining to claim that one exact combined passage existed. The project illustrates how conversational UX, carefully scoped instructions and historical source material can be combined to create an educational digital-humanities experience without presenting premodern theory as present-day clinical guidance.

16. Short Portfolio Card

Classical Medicine — Educational Encyclopedia

Custom GPT · Historical Research · Knowledge-Configured AI

An interactive Persian educational encyclopedia based on the five-volume Canon of Medicine, using staged case confirmation, thematic historical explanations, structured source identifiers and explicit non-clinical boundaries.

17. CV / Professional Summary

Classical Medicine — Educational Encyclopedia | Independent AI Product (2026)

Designed and published a Persian-language, knowledge-configured Custom GPT using Avicenna's five-volume Canon of Medicine. Developed a structured conversational workflow for hypothetical case studies, including intake, explicit summary confirmation, source-oriented thematic analysis, structured citation rules, uncertainty disclosure and non-clinical scope controls. Documented and evaluated a representative end-to-end interaction, and prepared the project for migration to a reusable plugin.

Compact CV version

Built a historical research Custom GPT around Avicenna's Canon, implementing confirmation-based conversational UX, source-reference rules, uncertainty handling and constrained educational outputs.

18. Interview Narrative

Situation: A large historical medical corpus was difficult to explore through ordinary keyword search, especially when a reader began with a complex fictional scenario rather than a known chapter.

Task: Build a Persian-language interactive reference that could connect questions to historical concepts without presenting itself as contemporary clinical expertise.

Action: Prepared a five-volume source collection, configured a Custom GPT, authored a multi-stage interaction contract, made source attribution and evidence gaps explicit, and added a confirmation checkpoint before analysis.

Result: Published the product and demonstrated an end-to-end case-study interaction in which the model retained scenario details, requested confirmation, discussed multiple relevant historical themes, produced chapter-style references and acknowledged the absence of an exact combined passage.

Boundary: This was configuration and application design using an existing LLM, not foundation-model training or a clinically validated decision-support system.

19. Suggested Metadata

Page title: Classical Medicine — Educational Encyclopedia | Conversational AI for Historical Research
Suggested slug: /classical-medicine-educational-encyclopedia
Category: AI Applications / Digital Humanities / Education
Tags: Custom GPT, Knowledge-Based AI, Avicenna, Canon of Medicine, History of Medicine, Persian, Conversational UX, Prompt Engineering, Source Grounding, Historical Research, Uncertainty Handling, Digital Humanities

Suggested social caption:

Can a centuries-old reference work become an interactive research experience? Classical Medicine — Educational Encyclopedia explores Avicenna's Canon of Medicine through structured Persian dialogue: fictional case intake, user confirmation, source-oriented historical interpretation and explicit evidence limitations. It is an educational project about historical medical thought—not a diagnostic or treatment tool.

20. Evidence and Attribution Notes

Directly provided / observed: public Custom GPT interface, creator-supplied Instructions, master.txt listed as its knowledge asset, conversation starters, and one complete confirmed example of the generated interpretation.

Creator description: the consolidated knowledge material represents the five volumes of the Canon of Medicine.

Not independently audited: edition completeness, every passage, correctness of each generated chapter reference, broader answer consistency, public usage scale, or clinical validity (which is outside the product's purpose).

This distinction is important for an accurate professional portfolio.

Final One-Line Statement

A Persian-language, knowledge-configured conversational encyclopedia that turns Avicenna's five-volume Canon of Medicine into a source-oriented historical learning experience with structured clarification, explicit confirmation, uncertainty handling and non-clinical output boundaries.