Editing Openai/691c1dba-9228-800f-8463-13b3a9006306 (section)

=== Fixes the API parsing, matches the real JSON shape returned by Ooba’s web UI. ===

Save as: hpl_runner_webui.py

<syntaxhighlight lang="python">#!/usr/bin/env python3
import json, hashlib, requests, time, os

API = os.environ.get("YAIY_LLM_API", "http://127.0.0.1:5000/api/v1/generate")
TIMEOUT = 90

def call_webui(prompt, max_tokens=512):
    payload = {
        "prompt": prompt,
        "max_new_tokens": max_tokens,
        "do_sample": True,
        "temperature": 0.7,
        "top_p": 0.9,
        "typical_p": 0.95
    }
    r = requests.post(API, json=payload, timeout=TIMEOUT)
    r.raise_for_status()
    data = r.json()

    # THIS IS THE REAL STRUCTURE:
    # {"results": [{"text": "model output..."}]}
    if "results" in data and len(data["results"]) > 0:
        return data["results"][0]["text"].strip()

    return str(data)

def carousel_ingot(lines, title="session"):
    text = "\n".join(lines).encode()
    digest = hashlib.sha256(text).hexdigest()[:32]
    syn = lines[-3:] if len(lines) > 3 else lines
    return {"title": title, "hash": digest, "synopsis": syn}

def run_manifest(manifest_path, context=""):
    manifest = json.load(open(manifest_path))
    logs = []
    for step in manifest["steps"]:
        p = f"""
--- 
Step ID: {step['id']}
Action: {step['do']}
Context: {context}
---
"""
        out = call_webui(p)
        logs.append(out)
        time.sleep(0.2)
    ingot = carousel_ingot(logs, title=manifest["title"])
    return {"logs": logs, "ingot": ingot}

if __name__ == "__main__":
    import argparse
    p = argparse.ArgumentParser()
    p.add_argument("manifest")
    p.add_argument("--context", default="Architect: Benjy")
    a = p.parse_args()
    print(json.dumps(run_manifest(a.manifest, a.context), indent=2))

</syntaxhighlight>

===== - Perfect for your GUI-based workflow =====
* Stable output extraction
* Supports all quantized GPTQ + GGML models loaded into webui
* Works with your dual-GPU setup