Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
from __future__ import annotations
import asyncio
import json
from types import SimpleNamespace
import pytest
from core.inference.tools import RAG_SOURCES_SENTINEL
from storage import research_runs_db as research_db
from storage import studio_db
STEPS = 12
REPORT = "## Findings\n\nThe manual covers it [Document: manual.pdf, p. 1]."
@pytest.fixture
def research_home(tmp_path, monkeypatch):
monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
monkeypatch.setattr(studio_db, "_schema_ready", set())
studio_db.upsert_chat_thread(
{
"id": "thread-1",
"title": "Research",
"modelType": "base",
"modelId": "local-model",
"createdAt": 1,
}
)
studio_db.upsert_chat_message(
{
"id": "user-1",
"threadId": "thread-1",
"role": "user",
"content": [{"type": "text", "text": "how does the pump work?"}],
"createdAt": 2,
}
)
return tmp_path
def _web_results(step: int) -> str:
return "\n\n---\n\n".join(
f"Title: Result {step}-{rank}\nURL: https://example.test/{step}/{rank}\n"
f"Snippet: web-{step}-{rank} " + "w" * 400
for rank in range(5)
)
def _passage(index: int) -> str:
return "d" * 800 + f" manual-fact-{index} " + "d" * 1_200
def _knowledge_base_results(step: int) -> str:
sources = [
{
"chunkId": f"chunk-{index}",
"documentId": "doc-1",
"filename": "manual.pdf",
"page": index + 1,
"text": _passage(index),
}
for index in range(step, step + 5)
]
text = "\n\n".join(source["text"] for source in sources)
return text + RAG_SOURCES_SENTINEL + json.dumps(sources)
def _synthesis_prompt(
monkeypatch,
knowledge_base,
context_length = None,
query = "pump",
max_sources = 100,
) -> str:
from core import research_runs as worker
supervisor = worker.ResearchSupervisor(SimpleNamespace(state = SimpleNamespace(server_port = 1)))
research_db.create_run(
run_id = "run-1",
owner_subject = "alice",
thread_id = "thread-1",
user_message_id = "user-1",
assistant_message_id = None,
config = {
"model": "local-model",
"inferenceRequest": {"model": "local-model"},
"ragScope": {"thread_id": "thread-1"},
"instructions": "",
"question": "how does the pump work?",
"budgets": {
"maxSteps": STEPS,
"maxSources": max_sources,
"modelTimeoutSeconds": 900,
"toolTimeoutSeconds": 10,
},
},
)
planned = research_db.set_plan(
"run-1",
{
"title": "Pump",
"steps": [
{"title": f"{query} {step}", "query": f"{query} {step}"} for step in range(STEPS)
],
},
)
research_db.approve("run-1", planned["planRevision"], planned["planHash"])
searches = {"web_search": 0, "search_knowledge_base": 0}
def fake_execute_tool(name, arguments, **kwargs):
step = searches[name]
searches[name] += 1
return _web_results(step) if name == "web_search" else knowledge_base(step)
prompts = []
async def fake_stream_completion(run, messages, **kwargs):
if kwargs.get("phase") == "synthesis":
prompts.append(messages[1]["content"])
return REPORT, "", "stop", None
return "not json", "", "stop", None
monkeypatch.setattr(worker, "_loaded_context_length", lambda _inf = None: context_length)
monkeypatch.setattr(worker, "execute_tool", fake_execute_tool)
monkeypatch.setattr(supervisor, "_stream_completion", fake_stream_completion)
asyncio.run(supervisor._process(research_db.claim_next(supervisor.worker_id)))
assert research_db.get_run("run-1")["status"] == "completed"
return prompts[-1]
def test_every_step_keeps_its_top_document_passage_in_the_report_prompt(research_home, monkeypatch):
prompt = _synthesis_prompt(monkeypatch, _knowledge_base_results)
assert all(f"manual-fact-{step}" in prompt for step in range(STEPS))
assert all(f"web-{step}-0" in prompt for step in range(STEPS))
def test_web_results_do_not_use_up_the_room_for_document_passages(research_home, monkeypatch):
prompt = _synthesis_prompt(monkeypatch, _knowledge_base_results, max_sources = 40)
assert all(f"manual-fact-{step}" in prompt for step in range(STEPS))
assert all(f"Chunk ID: chunk-{step}\n" in prompt for step in range(STEPS))
assert len(research_db.get_run("run-1")["sources"]) == 40
def test_small_context_keeps_passages_when_both_catalogs_fill(research_home, monkeypatch):
def distinct_chunks(step):
sources = [
{
"chunkId": f"{index:08d}-1111-4222-8333-444455556666",
"documentId": "9b2f6c1e-7d4a-4f0b-9c3e-2a1b5d6e7f80",
"filename": "manual.pdf",
"page": index + 1,
"text": _passage(index),
}
for index in range(step * 5, step * 5 + 5)
]
return (
"\n\n".join(source["text"] for source in sources)
+ RAG_SOURCES_SENTINEL
+ json.dumps(sources)
)
prompt = _synthesis_prompt(monkeypatch, distinct_chunks, context_length = 8_192, max_sources = 40)
assert all(f"web-{step}-0" in prompt for step in range(STEPS))
assert "https://example.test/0/0" in prompt
assert "Chunk ID: 00000000-1111-4222-8333-444455556666" in prompt
def test_web_results_keep_their_room_when_the_documents_match_nothing(research_home, monkeypatch):
prompt = _synthesis_prompt(
monkeypatch, lambda _step: "No matching chunks were found in the knowledge base."
)
assert all(f"web-{step}-{rank}" in prompt for step in range(STEPS) for rank in range(5))
def test_small_context_keeps_web_results_when_the_documents_match_nothing(
research_home, monkeypatch
):
prompt = _synthesis_prompt(
monkeypatch,
lambda _step: "No matching chunks were found in the knowledge base.",
context_length = 8_192,
query = "centrifugal pump cooling data sheet maintenance interval seal replacement guide",
)
assert all(f"web-{step}-0" in prompt for step in range(STEPS))
def test_web_only_catalog_is_not_trimmed(research_home, monkeypatch):
prompt = _synthesis_prompt(
monkeypatch,
lambda _step: "No matching chunks were found in the knowledge base.",
context_length = 4_096,
)
catalog = prompt.split("<source_catalog>\n", 1)[1].split("</source_catalog>", 1)[0]
assert catalog.count("URL: ") == len(research_db.get_run("run-1")["sources"])