README.md

metadata

title: QUEST
emoji: 🔎
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.29.0
app_file: app.py
pinned: false

DeepResearch Space

An interactive Hugging Face Space for a Quest DeepResearch agent. The app can either talk to osunlp/QUEST-35B (our own fine-tuned research model, routed through a private HF Inference Endpoint) or fall back to open-weights models through the shared HF Inference API.

Supported tools:

• search (DuckDuckGo, multi-query)
• visit (HTTP fetch + text extraction, multi-URL)
• lightweight research-state summary to cut repeated work
• <answer> extraction for the final response

---

1) Use our own osunlp/QUEST-35B model (recommended)

Because the model is private during the beta, it is not on the free Inference API. You host it yourself on a dedicated HF Inference Endpoint (pay-as-you-go, scale-to-zero), and point this Space at it.

1a) Create the endpoint once

1. Open https://ui.endpoints.huggingface.co/ and click "New endpoint".
2. Model repository: osunlp/QUEST-35B (use a token with access).
3. Hardware: 1x Nvidia L4 (24GB) is usually the sweet spot for a 35B model. Nvidia T4 small (16GB) works too and is cheaper.
4. Advanced → Container Type: keep Text Generation Inference (TGI) or pick vLLM. Both expose an OpenAI-compatible /v1/ route.
5. Autoscaling → Scale-to-Zero: enable it so you only pay when the endpoint is serving traffic.
6. Hit Create endpoint. After ~1–2 minutes it turns Running and shows a base URL like https://abcdef.us-east-1.aws.endpoints.huggingface.cloud.

1b) Tell the Space how to reach it

In this Space's Settings → Secrets / Variables:

Name	Value	Why
HF_TOKEN	your personal HF token with read access to osunlp/QUEST-35B	pulls private weights & authenticates the endpoint call
QUEST_BASE_URL	the endpoint URL ending with /v1/ (e.g. https://abcdef.us-east-1.aws.endpoints.huggingface.cloud/v1/)	tells the app to route chat completions to your endpoint
QUEST_ENDPOINT_MODEL	tgi (default; set to the original repo id osunlp/QUEST-35B if you deployed with vLLM)	some containers need the exact model name
DEFAULT_MODEL	osunlp/QUEST-35B	preselects the right option in the UI

Click Restart this Space. The Model dropdown now shows osunlp/QUEST-35B at the top; selecting it routes requests through your endpoint.

Cost reality-check: on a 1× L4 at $0.80/hr with Scale-to-Zero, a small internal beta (a handful of testers, dozens of queries per day) typically stays under $100/month. You can stop the endpoint manually from the UI any time to freeze costs.

---

2) Fallback: free open-weights models

If you just want to try the UI without spinning up an endpoint, pick any of these in the dropdown. They run through the shared HF Inference API.

• Qwen/Qwen3-8B
• google/gemma-3-12b-it
• deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
• Qwen/Qwen2.5-7B-Instruct
• meta-llama/Llama-3.1-8B-Instruct

Only HF_TOKEN is required for this path.

---

3) Share the beta with org members (without paying for Team)

Option A (simplest, $0 for access, Space Hardware stays on free CPU):

1. Keep the Space under your personal account.
2. Settings → Visibility → Private.
3. Settings → Collaborators → add each tester by HF username.
4. Endpoint lives under your personal namespace too, so the bill goes to your personal payment method (you can expense invoices from https://huggingface.co/settings/billing).

Option B (org-level billing): upgrade the organization to a Team plan and recreate both the Space and the endpoint under the org namespace.

---

4) Local development

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export HF_TOKEN=...                      # required
export QUEST_BASE_URL=https://.../v1/    # optional; only if testing against the endpoint
python app.py

---

5) Architecture notes

• app.py uses huggingface_hub.InferenceClient(base_url=QUEST_BASE_URL, ...) for the private-endpoint path and the same client without base_url for the shared API path.
• The system prompt matches the schema QUEST-35B was trained on (array-based search / visit with an explicit goal), so the private model stays in-distribution. The open-weights fallbacks also follow the same schema.
• Visited URLs and search queries are cached in-process so repeated tool calls don't re-hit the network.
• <answer>...</answer> terminates the ReAct loop.