| import sys, json, re, argparse, gradio as gr |
| from crop_tools import execute |
| ap = argparse.ArgumentParser(); ap.add_argument("--backend", default="ollama", choices=["ollama","hf"]) |
| ap.add_argument("--model", default="commodities-crop-analyst"); ap.add_argument("--adapter", default="."); ap.add_argument("--base", default="Qwen/Qwen2.5-7B-Instruct") |
| args = ap.parse_args() |
| import datetime |
| SYSTEM = (open("system_prompt.txt").read() if args.backend == "ollama" else [json.loads(l) for l in open("data/val_v2.jsonl")][0]["messages"][0]["content"]) + f"\nToday's date is {datetime.date.today().isoformat()}. Resolve relative dates (last year, a year ago, last month) from it. Price data runs 2017-01-03 to {open('data/futures_daily.parquet','rb') and __import__('pandas').read_parquet('data/futures_daily.parquet').date.max()}." |
|
|
| if args.backend == "hf": |
| import torch |
| from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig |
| from peft import PeftModel |
| tok = AutoTokenizer.from_pretrained(args.base) |
| model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(args.base, quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16), device_map="auto"), args.adapter).eval() |
| def gen(msgs, n=300): |
| ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)["input_ids"].to(model.device) |
| return tok.decode(model.generate(ids, max_new_tokens=n, do_sample=False)[0][ids.shape[1]:], skip_special_tokens=True).strip() |
| else: |
| import requests |
| def gen(msgs, n=300): |
| r = requests.post("http://localhost:11434/api/chat", json={"model": args.model, "messages": msgs, "stream": False, "options": {"temperature": 0, "num_predict": n}}, timeout=300) |
| r.raise_for_status(); return r.json()["message"]["content"].strip() |
|
|
| def answer(user, history): |
| msgs = [{"role":"system","content":SYSTEM}] |
| for h in history[-6:]: |
| msgs.append({"role":h["role"],"content":h["content"] if isinstance(h["content"], str) else "[image]"}) |
| msgs.append({"role":"user","content":user}) |
| first = gen(msgs, 160) |
| m = re.search(r"<tool_call>(.*?)</tool_call>", first, re.S) |
| if not m: return first, None, None |
| try: call = json.loads(m.group(1)) |
| except Exception: return "I produced a malformed tool call; please rephrase.", None, first |
| res = execute(call) |
| msgs += [{"role":"assistant","content":first}, {"role":"tool","content":json.dumps(res)}] |
| final = gen(msgs, 300) |
| img = res.get("image") if isinstance(res, dict) else None |
| return final, img, json.dumps({"call": call, "result": res}, indent=1, default=str) |
|
|
| def chat(user, history): |
| text, img, trace = answer(user, history) |
| out = [{"role":"assistant","content":text}] |
| if img: out.append({"role":"assistant","content":{"path": img}}) |
| if trace: out.append({"role":"assistant","content":"```json\n" + trace[:1500] + "\n```", "metadata":{"title":"tool call"}}) |
| return out |
|
|
| gr.ChatInterface(chat, title="Commodities crop analyst (edoggy)", |
| description="Answers from USDA CDL area, census, IBGE yields and futures data via tools. Forecasts are statistical baselines with error bars. Not advice.", |
| examples=["How much corn did Iowa have in 2023?", "Which Louisiana parish has the most sugarcane?", "Forecast sugar 60 days out", |
| "Show me the sugarcane map for Iberia, LA", "Chart orange juice since 2022", "Is Louisiana cane expanding next year?"] |
| ).launch(server_name="0.0.0.0", server_port=7860) |
|
|
