Provider Control

AI Configuration

Manage large language models, service endpoints, and specialized grading tasks.

Service Connections

Configure provider endpoints and connection status.

Active

OpenAI Connection

External

Openai integration for grading workflows.

Connection Status Not Tested

Active

Anthropic Claude Connection

External

Claude integration for grading workflows.

Connection Status Not Tested

Active

Google Gemini Connection

External

Gemini integration for grading workflows.

Connection Status Not Tested

Active

DeepSeek Connection

External

Deepseek integration for grading workflows.

Connection Status Not Tested

Disabled

LM Studio Connection

Trusted

Lm-studio integration for grading workflows.

Connection Status Disabled

LM Studio access works best when this app runs locally (HTTP) with LM Studio CORS enabled.

Disabled

Ollama Connection

External

Ollama integration for grading workflows.

Connection Status Disabled

Active

Chrome Built-in AI Connection

Trusted

Chrome-ai integration for grading workflows.

Connection Status Not Tested

Add Provider

Connect your own API key or custom endpoint.

PII & Processing

Ensure student credentials are protected before any provider processing.

Require internal processing

Trusted, local, and internal processing only.

Allow external providers

Allow processing with non-local providers when policy permits.

Redact student names

Replace detected student name values before outbound requests.

Redact student IDs

Mask student IDs and number fields for external processing.

Redact grade values

Hide grade-related values from outbound payloads.

Model Assignments

Define which model handles specific academic workloads.

Task Category	Model	Provider
ChatGeneral-purpose chat used in the grading assistant.	Primary gpt-5-mini	Fallback OpenAI
Complex ReasoningMatches extracted student identity against the roster. Prefers a stronger model.	Primary gpt-5-mini	Fallback OpenAI
Identity OCRCleans up OCR noise from handwritten name/ID fields before roster matching.	Primary gpt-5-mini	Fallback OpenAI
Split DetectionDetects exam page boundaries when splitting a scanned PDF.	Primary gpt-5-mini	Fallback OpenAI
Zone DetectionLocates answer zones on a scanned exam page.	Primary gpt-5-mini	Fallback OpenAI
Identity HintingExtracts student name / ID field locations from an exam template.	Primary gpt-5-mini	Fallback OpenAI
Code OCRTranscribes handwritten code answers into text.	Primary gpt-5-mini	Fallback OpenAI
Answer ParsingParses a raw OCR transcript into structured answer objects.	Primary gpt-5-mini	Fallback OpenAI
Answer GradingScores a parsed answer against the solution.	Primary gpt-5-mini	Fallback OpenAI
Solution InferenceInfers expected answers from an unmarked template scan.	Primary gpt-5-mini	Fallback OpenAI

PDF Split Strategy

Choose how scanned exam batches are split into per-student PDFs.

LLM

Use the AI provider configured above to detect student boundaries.

Deterministic OCR + fingerprinting

Browser-only pipeline using PDF.js, Tesseract, dHash and token clustering. No network calls.

Zone Detection Strategy

Choose how exam zones (name, ID, question regions) are detected. Solution inference fills in points and answer keys later.

LLM OCR

Use the AI provider to detect zones from one reference page. One template applied to all students.

LLM OCR (per student)

Run the AI vision model on each student's pages individually. Higher accuracy per student but slower and uses more AI credits.

OCR Template

Tesseract OCR on one reference page. Finds the PII page and precise identity-field zones locally.

OCR Eager (per student)

Per-student zones with ink-extent refinement for answers plus local PII page and identity-field detection.

Student Matching Strategy

Choose how detected identity zones are matched against the roster.

Deterministic OCR

Use local Tesseract OCR on identity zones and deterministic roster scoring. No LLM calls.

TrOCR Fuzzy Matching

Use local TrOCR handwriting recognition on identity zones and fuzzy roster ranking. No LLM calls.

LLM Matching

Use OCR-extracted identity text with LLM normalization and roster matching fallback.

LLM — Single Step

One LLM call per submission: corrects OCR noise and matches against the full roster in a single prompt. Best when using a strong reasoning model.

Answer Parsing

Optional step that digitizes each student answer from the scanned PDF and persists a typed representation. When a verified parsed answer exists for a cell, autograding uses it as text (cheaper, deterministic for MCQ) instead of re-running vision on the same image. Manage which model handles Answer Parsing in Model Assignments above.

Strategy

LLM (vision)

Cropped student answer image is sent to the configured Answer Parsing model with type-aware prompts (text, code, MCQ, MSQ, numeric).

Deterministic coming soon

Tesseract / TrOCR-driven parsing without an LLM.

Autograding

Configure how student submissions are auto-graded. The LLM picks one category per question; graders can override any pick from the autograding page. Manage which model handles Answer Grading in Model Assignments above.

Strategy

LLM (vision)

Cropped student answer image is sent to the configured Answer Grading model along with the inferred reference solution.

Deterministic coming soon

Rule-based scoring (e.g. exact-match for MCQ keys) without an LLM.

Grading Categories

LabelScore multiplier (0–1)

Fallback category (used when LLM output is unparseable):

Leniency Controls (Forwarded To LLM)

These controls guide grading tolerance. They are appended to the autograding prompt.

Grading style:

Ignore minor syntax mistakes when intent/correctness is clearIgnore identifier underscore prefix differences (example: `_a` vs `a`)Ignore formatting-only differences (spacing, line breaks, superficial style)Additional grading guidance:

Storage & Backup

Monitor browser storage usage, export your data, or clear stored records by category.

Loading storage info…

Local browser settings mode.