title	Chat Cooking
emoji	🍳
colorFrom	yellow
colorTo	red
sdk	gradio
sdk_version	6.15.2
python_version	3.11
app_file	app.py
pinned	false

Chat Cooking

A cooking assistant chatbot powered by RAG (Retrieval-Augmented Generation). Ask about recipes, ingredients, or cooking techniques — it semantically retrieves relevant recipes from a dataset of 1,090 dishes and streams a grounded response via the HuggingFace Inference API.

Live demo: huggingface.co/spaces/kawas8516/chat-cooking

---

How it works

User message
      |
      v
is_recipe_query()          keyword + word-boundary classifier
      |
   Yes |  No ──────────────────────────────────────────> LLM (general answer)
      |
      v
retrieve()                 FAISS IndexFlatIP over sentence-transformers embeddings
      |                    (all-MiniLM-L6-v2, cosine similarity, top-k results)
      v
build_messages()           system prompt + retrieved context + conversation history
      |
      v
stream_response()          HuggingFace Inference API  (Qwen2.5-7B-Instruct, streaming)
      |
      v
Gradio UI                  chat panel (left) | retrieved recipes panel (right)

Key design decisions

Decision	Rationale
FAISS IndexFlatIP over normalized vectors	Exact cosine search, no approximation error on a 1K corpus
all-MiniLM-L6-v2 embedder	384-dim, fast on CPU, strong semantic recall for food queries
Keyword classifier before retrieval	Avoids embedding cost and noisy retrieval on non-food questions
Retrieved context injected as a second system message	Keeps it clearly separated from user history for the LLM
Name + ingredients and directions sent as context	The model is told not to invent recipes, so it needs the actual steps to ground on
SCORE_THRESHOLD gate on retrieval	Drops low-relevance matches so off-topic queries get a general answer, not spurious recipes
Retrieval query includes the previous user turn	Follow-ups ("make it vegetarian") retrieve in the context of the original dish
Index built at startup if missing	Binary files rejected by HF git; CSV stays in repo, index is ephemeral

---

Evaluation

Measured on the 1,090-recipe dataset with python eval.py:

Metric	Score
Classifier accuracy (14 labelled queries)	100%
Retrieval Hit@3 (12 query/recipe pairs)	92%
Retrieval MRR	0.861
Top-1 retrieval score (mean / min)	0.720 / 0.592
Faithfulness (6 cases, LLM-as-judge, --faithfulness)	83%

The one retrieval miss ("chocolate cherry dessert") is a data-coverage gap — no exact match exists in the 1,090-recipe subset. MRR of 0.861 means the correct recipe is ranked #1 on average. Mean top-1 similarity of 0.720 (min 0.592) sits comfortably above the SCORE_THRESHOLD of 0.3, so relevant queries always clear the relevance gate.

Faithfulness is a small (6-case), non-deterministic LLM-judge check; the one miss adds a couple of plausible details (lemon juice, paprika) absent from the retrieved fish recipe.

Run evaluation yourself:

python eval.py                  # classifier + retrieval
python eval.py --faithfulness   # + LLM faithfulness judge (uses tokens)
python eval.py --k 5            # evaluate Hit@5 instead
python eval.py --json out.json  # save results to JSON

---

Project structure

chat-cooking/
├── app.py              # Gradio Blocks UI
├── rag.py              # retrieval: embedder, FAISS index, is_recipe_query()
├── llm.py              # build_messages(), stream_response()
├── config.py           # env-based configuration (python-dotenv)
├── build_index.py      # build data/recipes.index from data/recipes.csv
├── eval.py             # evaluation suite (classifier + retrieval + faithfulness)
├── data/
│   └── recipes.csv     # 1,090 recipes (name, ingredients, directions, url)
├── tests/
│   ├── test_rag.py     # 12 unit tests for retrieval and classifier
│   └── test_llm.py     # 11 unit tests for message building and streaming
└── .github/workflows/
    └── ci.yml          # ruff + pytest on Python 3.11

---

Quickstart (local)

git clone https://github.com/kawas8516/chat-cooking
cd chat-cooking
pip install -r requirements.txt
cp .env.example .env        # add your HF_TOKEN
python build_index.py       # embeds 1090 recipes, writes data/recipes.index (~15s)
python app.py               # open http://localhost:7860

Environment variables

Variable	Default	Description
HF_TOKEN	required	HuggingFace token (Inference API access)
MODEL_ID	Qwen/Qwen2.5-7B-Instruct	Generation model
EMBED_MODEL_ID	all-MiniLM-L6-v2	Embedding model for retrieval
TOP_K	3	Number of recipes retrieved per query
SCORE_THRESHOLD	0.3	Min cosine similarity for a recipe to be used as context
MAX_NEW_TOKENS	512	Max tokens per LLM response
TEMPERATURE	0.7	Sampling temperature

Alternative models (set MODEL_ID to swap with no code changes):

• meta-llama/Llama-3.2-3B-Instruct — lighter, faster
• mistralai/Mistral-7B-Instruct-v0.3 — strong instruction following

---

Deploy to Hugging Face Spaces

1. Fork this repo
2. Create a new Space at huggingface.co/new-space — SDK: Gradio
3. Push to the Space remote: git push hf-space main
4. Add HF_TOKEN under Settings > Secrets
5. The Space auto-builds the index on first boot — no manual step needed

---

Development

pip install -r requirements-dev.txt
python -m pytest          # 23 tests
ruff check .              # lint
python eval.py            # evaluation

CI runs on every push to main via GitHub Actions (.github/workflows/ci.yml).

---

Stack

• UI — Gradio Blocks
• Retrieval — FAISS + sentence-transformers
• Generation — HuggingFace Inference API (streaming)
• Data — 1,090 recipes scraped from Allrecipes