LM Studio Subagents

lms load qwen2.5-7b-instruct --identifier local-worker --context-length 16384
curl http://localhost:1234/v1/models

import os
import requests

LM_STUDIO_URL = "http://localhost:1234"
TOKEN = os.environ.get("LM_API_TOKEN")  # only needed when server auth is enabled
HEADERS = {"Authorization": f"Bearer {TOKEN}"} if TOKEN else {}
MODEL = os.environ.get("LM_MODEL", "local-worker")

def _chat(**body) -> dict:
    body.setdefault("model", MODEL)
    body.setdefault("max_output_tokens", 4096)
    body.setdefault("reasoning", "off")   # reasoning tokens eat the output budget
    body.setdefault("store", True)        # required for previous_response_id
    r = requests.post(f"{LM_STUDIO_URL}/api/v1/chat", headers=HEADERS, json=body,
                      timeout=900)        # local models are slow; do not time out early
    r.raise_for_status()
    data = r.json()
    message = next(item for item in data["output"] if item["type"] == "message")
    return {"text": message["content"], "response_id": data.get("response_id"),
            "stats": data.get("stats", {})}

def ask_local(prompt: str, system: str = "You are a helpful assistant.") -> dict:
    return _chat(input=prompt, system_prompt=system, temperature=0.3,
                 context_length=32768)   # match the loaded model's configured context

def continue_local(response_id: str) -> dict:
    """Resume a cut-off answer without resending the source text."""
    return _chat(previous_response_id=response_id,
                 input="Continue exactly where you stopped. Do not repeat prior text.")

result = ask_local("Summarize this text in 2 sentences: ...")
print(result["text"])
# If stats["total_output_tokens"] reached max_output_tokens, the answer was cut off; extend it:
# result = continue_local(result["response_id"])

from openai import OpenAI

# Point to local LM Studio server
client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio",  # any non-empty string; only checked when auth is on
)

def ask_local(prompt: str, system: str = "You are a helpful assistant.") -> str:
    response = client.chat.completions.create(
        model="local-worker",  # the identifier shown by /v1/models
        messages=[
            {"role": "system", "content": system},
            {"role": "user", "content": prompt},
        ],
        temperature=0.3,
        max_tokens=4096,  # generous; check finish_reason below
    )
    if response.choices[0].finish_reason == "length":
        raise RuntimeError("LM Studio truncated the response; use the native API continuation flow.")
    return response.choices[0].message.content

# Example usage
result = ask_local("Summarize this text in 2 sentences: ...")
print(result)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

class LocalSubagent:
    def __init__(self, system_prompt: str, temperature: float = 0.2):
        self.system_prompt = system_prompt
        self.temperature = temperature

    def run(self, user_input: str) -> str:
        response = client.chat.completions.create(
            model="local-worker",
            messages=[
                {"role": "system", "content": self.system_prompt},
                {"role": "user", "content": user_input},
            ],
            temperature=self.temperature,
            max_tokens=4096,  # generous; check finish_reason below
        )
        if response.choices[0].finish_reason == "length":
            raise RuntimeError("Response truncated; retry through the native API continuation flow.")
        return response.choices[0].message.content

# Define specialized subagents
summarizer = LocalSubagent(
    system_prompt="You are a summarization expert. Produce concise 2-3 sentence summaries."
)

classifier = LocalSubagent(
    system_prompt="Classify the input into one of these categories: billing, technical, general, urgent. Respond with only the category name.",
    temperature=0.0,
)

extractor = LocalSubagent(
    system_prompt="Extract all named entities (people, organizations, dates, amounts) from the text. Return as JSON.",
    temperature=0.0,
)

# Use the subagents
summary = summarizer.run("Long document text here...")
category = classifier.run("I can't log into my account and I need to submit a report by EOD")
entities = extractor.run("John Smith signed a $50,000 contract with Acme Corp on March 15, 2025")

import asyncio
from openai import AsyncOpenAI

client = AsyncOpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

async def process_batch(items: list[str], system_prompt: str, max_concurrent: int = 4,
                        max_tokens: int = 4096) -> list[dict]:
    semaphore = asyncio.Semaphore(max_concurrent)

    async def process_one(text: str) -> dict:
        async with semaphore:
            response = await client.chat.completions.create(
                model="local-worker",
                messages=[
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": text},
                ],
                temperature=0.2,
                max_tokens=max_tokens,  # generous but finite: unbounded batches can run for hours
            )
            choice = response.choices[0]
            # Flag truncation per item instead of raising — one bad item must not
            # discard the other 99 results.
            return {"text": choice.message.content, "truncated": choice.finish_reason == "length"}

    tasks = [process_one(item) for item in items]
    results = await asyncio.gather(*tasks, return_exceptions=True)
    return [r if isinstance(r, dict) else {"text": None, "error": str(r)} for r in results]

# Batch summarize 100 documents
documents = ["doc1 text...", "doc2 text...", ...]  # 100 documents
summaries = asyncio.run(process_batch(
    documents,
    system_prompt="Summarize in 2 sentences.",
    max_concurrent=2,  # keep low on a single local GPU
))
incomplete = [i for i, r in enumerate(summaries) if r.get("truncated") or r.get("error")]

def smart_route(task_type: str, text: str) -> str:
    """Route tasks between local and cloud models."""
    local_tasks = {"summarize", "classify", "extract_simple", "reformat"}

    if task_type in local_tasks:
        return ask_local(text)  # Free, local inference
    else:
        return ask_cloud(text)  # Paid, cloud API

tickets = load_tickets_from_csv("tickets.csv")
summaries = asyncio.run(process_batch(
    [t["description"] for t in tickets],
    system_prompt="Summarize this support ticket in one sentence. Include the main issue and any resolution.",
    max_concurrent=2,
))
# Re-run any item where r["truncated"] is True through the native continuation flow.
# Cost: $0 (vs ~$15 with GPT-4)

classifier = LocalSubagent(
    system_prompt="Classify this email into exactly one category: sales, support, spam, internal. Reply with only the category.",
    temperature=0.0,
)
for email in emails:
    category = classifier.run(email["subject"] + "\n" + email["body"])
    email["category"] = category.strip().lower()

extractor = LocalSubagent(
    system_prompt='Extract fields from the contract as JSON: {"parties": [], "date": "", "amount": "", "term": ""}',
    temperature=0.0,
)
for doc in contracts:
    data = extractor.run(doc["text"])
    print(f"{doc['filename']}: {data}")