Recipes

Zero-infrastructure RAG cookbook example using ChromaDB and Ollama — no Docker, no external services.

What this demonstrates

This example shows Orchid running with zero infrastructure: no Docker, no Qdrant, no external API keys. The vector store is ChromaDB, an embedded, on-disk database that persists vectors at ~/.orchid/chroma/. The LLM and embedding model both run locally through Ollama.

Two agents — cookbook and mealplanner — draw on a shared recipe corpus seeded at startup. cookbook answers ingredient questions and suggests substitutions; mealplanner builds dietary-aware weekly meal plans. Both use simple dense retrieval with reformulation so natural-language queries match the recipe corpus regardless of phrasing.

Run it

# From the repo root
pip install -e orchid -e orchid-cli

# Send a recipe question
orchid chat send "What can I make with chicken and rice?" \
  --config examples/recipes/orchid.yml

# Ask for a meal plan
orchid chat send "Plan a vegan dinner for Monday" \
  --agent mealplanner \
  --config examples/recipes/orchid.yml

# Interactive session
orchid chat interactive --config examples/recipes/orchid.yml

Ollama must be running locally with llama3.2 and nomic-embed-text pulled.

Configuration walkthrough

orchid.yml

llm:
model: ollama/llama3.2

auth:
dev_bypass: true

rag:
vector_backend: chroma          # on-disk, no Docker needed
embedding_model: ollama/nomic-embed-text   # 768-d, runs locally

storage:
class: orchid_ai.persistence.sqlite.OrchidSQLiteChatStorage
dsn: ~/.orchid/recipes.db

startup:
hook: examples.recipes.hooks.startup.seed_recipes

The rag.vector_backend: chroma key is what makes this example zero-infra — ChromaDB stores vectors as files under ~/.orchid/chroma/ via its PersistentClient. No Docker container, no Qdrant URL. The startup.hook seeds 8 recipes at bootstrap so agents have material to retrieve without manual indexing.

agents.yaml

version: "1"

defaults:
llm:
  model: ollama/llama3.2
  temperature: 0.2

agents:
cookbook:
  description: >
    General cooking and recipe assistant.  Answers questions about
    ingredients, techniques, substitutions.
  prompt: |
    You are a helpful chef assisting with cooking questions.
    Answer using the retrieved recipe documents.
  rag:
    namespace: recipes
    k: 5
    retrieval:
      strategy: simple
      query_transformers: [reformulate]

mealplanner:
  description: >
    Meal-planning assistant that considers dietary restrictions
    and available ingredients.
  prompt: |
    You are a meal-planning assistant.  Build meal plans from the
    retrieved recipe documents.
  rag:
    namespace: recipes
    k: 8
    retrieval:
      strategy: simple
      query_transformers: [reformulate]

The cookbook agent lives in agents/cookbook.md:

---
description: "General cooking and recipe assistant."
rag:
  namespace: recipes
  k: 5
  retrieval:
    strategy: simple
    query_transformers: [reformulate]
---

You are a helpful chef assisting with cooking questions.
Answer using the retrieved recipe documents.

Both agents share the recipes namespace. The query_transformers: [reformulate] flag smooths out natural-language questions — a user asking "What can I cook with stuff in my fridge?" gets useful results even when the query structure is loose.

What to look for

  • vector_backend: chroma — the single YAML key that swaps the entire RAG backend from Qdrant to ChromaDB. No code changes, no config restructure.
  • Startup hook seeding — hooks/startup.py uses reader.upsert() to populate 8 recipes into ChromaDB at bootstrap. Each recipe carries metadata (cuisine, course, prep time, dietary tags) available for filtered retrieval.
  • Zero Docker — ChromaDB is an embedded database. ~/.orchid/chroma/ is just files on disk. You can inspect them or delete them to start fresh.
  • Ollama all the way — both the LLM (llama3.2) and embeddings (nomic-embed-text) run locally. No OpenAI or Gemini API key required.
  • Metadata filtering — recipes are tagged with cuisine, course, dietary, prep_time_min, and difficulty. Query with metadata_filters: {"cuisine": "italian"} to narrow results.

Related concepts

  • Hierarchical RAG — scope hierarchy, retrieval strategies, metadata filtering
  • Embeddings — supported models, dimension differences
  • Agents — YAML-driven agent configuration