North Robot - Cytation 5 uSDL

All code written by Owen Melville with contributions by Ilya Yakavets, Monique Ngan and ChatGPT.

File List:

• Locator.py: Contains a list of physical locations used by the Robot

• North_Safe.py Contains the reusable code for the control of the North Robot and the North Track which goes between the Cytation 5 and the North Robot. This is the most in-depth file.

• biotek.py Contains the code for control of the Cytation 5. Ilya has a newer version with more capabilities.

• master_usdl_coordinator.py Contains reusable workflow code that uses multiple instruments: Cytation 5, North Robot & Track, Photoreactor

• north_gui.py Mostly written by chatGPT, intended as a base for a resuable gui for different workflows

• photoreactor_controller Controls the photoreactor via Raspberry Pi Pico. reactor_test.py is the local Pico program.

• requirements.txt What packages need installation to use all this?

• slack_agent.txt Controls messages sent to slack from the robot. Mostly written by ChatGPT.

Folder List:

• analysis : Contains programs that analyze data from the robot to produce a result usable by the recommenders

• photo-reactor : Contains the RPi program for the photoreactor.

• recommenders: Contains programs (eg using Baybe) to recommend conditions

• status: Contains transferrable data structures (eg vial status). In the long run these will represent objects with states that transfer between modules

• tests: Contains short test or commonly used programs for the setup.

• workflows: Contains active workflows that are designed for the setup, such as the color matching workflow.

Sample Workflow:

The sample workflow contained in this video is labelled and contains the programming elements in the sample_workflow.py program in the "workflows" directory.

Literature Relevance Labeling (LLM Assisted)

Pipeline components live under research/literature_search/:

1. Fetch + extract + score produce scored_candidates.csv and parsed JSONL.
2. Gating selects top percentile candidates; exploration sampling can add mid/low segments.
3. prompt_preview.py (already run) renders offline prompts for inspection.
4. llm_label.py sends prompts to an OpenAI-compatible model and writes JSONL labels.

.env Setup

Create a .env file at repo root (already git-ignored via *.env) with:

OPENAI_API_KEY=sk-...
# Optional self-host / proxy endpoint:
# OPENAI_BASE_URL=https://your-proxy.example.com

Install Dependencies

Ensure openai and python-dotenv were installed (added to requirements.txt).

Dry Run (Schema / Flow Check)

python research/literature_search/scripts/llm_label.py --input research/literature_search/data/prompt_preview_new.jsonl --output research/literature_search/data/llm_labels_dry.jsonl --dry-run

Real Labeling Run

python research/literature_search/scripts/llm_label.py --input research/literature_search/data/prompt_preview_new.jsonl --output research/literature_search/data/llm_labels.jsonl --model gpt-4o-mini --rate-limit-per-min 40

--rate-limit-per-min is a client-side throttle; adjust per your quota. Output lines contain:

{"id": "...", "model_output": { ... schema ... }}

Next Steps (Optional)

• Aggregate labels and fit calibration (logistic) from axis vector -> relevance probability.
• Consider renaming device_penalty axis after initial calibration so all axes share the same directionality (higher=better) before probability fitting.