vilsonkhchoian/promting_project_iad

Purpose

Essay Checker Bot is a Telegram workflow for evaluating English exam essays against HSE-style criteria. A student submits an essay task and essay text; the system asks GigaChat for a 0–10 score plus feedback. Core runtime logic is in bot.py; provider and prompt settings are centralized in config.py.

Runtime Entry Point

bot.py is the executable application boundary. It receives Telegram updates, manages the multi-message checking interaction, invokes the LLM, and returns the generated assessment. /start begins a new evaluation and /cancel clears an in-progress one.

Paths: bot.py, config.py

Conversation-State Pipeline

The bot implements a sequential collection flow:

1. User starts a session.

2. Bot receives and retains the essay assignment.

3. Bot receives the essay body, including long responses.

4. The combined task/essay payload is submitted for evaluation.

5. The resulting score and comments are delivered back through Telegram.

This stateful chat boundary separates untrusted user input from the model request construction.

Paths: bot.py

Prompted Evaluation Layer

Evaluation behavior is driven by a prompt template and model-generation parameters: endpoint URLs, model selection, temperature, maximum tokens, and credentials. The intended rubric covers task response, cohesion, vocabulary/register, and grammar. Prompt design includes an injection-resistant variant and was empirically compared before being used in the bot.

Paths: config.py, Промпты.pdf, bot.py

External-Service Boundaries

Telegram is the inbound/outbound messaging transport; its bot token is configured outside application logic. GigaChat is the production inference provider, accessed through configured API endpoints and API credentials. The application depends on network availability and provider response formatting when producing a user-visible grade.

Paths: bot.py, config.py, requirements.txt

Offline Model and Prompt Evaluation

The notebook is the experimentation and validation environment rather than the serving path. It runs essay checks against GigaChat and OpenRouter-accessed models, retries poor responses, and calculates evaluation metrics. Its findings are persisted as spreadsheets: detailed per-essay/model/prompt results and aggregate ranking metrics such as MAE, RMSE, Spearman, and QWK.

Paths: essay_prompt_project.ipynb, model_scores.xlsx, metrics_summary.xlsx

Supporting Project Artifacts

Project documentation records the chosen criteria, candidate models, prompt variants, experimentation methodology, metrics, and prompt-injection considerations. These artifacts define the quality-governance context for future changes to the prompt or underlying model, but are not application runtime dependencies.

Paths: Проект ПЗАД.pdf, Промпты.pdf, README.md