Education Studio

Four-agent teaching workflow with concept extraction, quiz generation, lesson planning, and multi-format export via PDF, DOCX, PPTX, and Markdown.

What this demonstrates

The Education Studio example turns uploaded course material into quizzes, lesson plans, and export-ready teaching packages. Four specialized agents — content-analyzer, quiz-generator, lesson-builder, and format-exporter — collaborate through three cross-agent skills (generate_quiz, generate_lesson, generate_full_package). The quiz-generator uses mini-agents to draft questions from multiple source sections in parallel. All agents share RAG-scoped namespaces for source material, generated content, and exports. A cron-driven weekly-quiz schedule and a chat-bound education.generate signal trigger event-driven generation without manual intervention.

Run it

pip install -e ./orchid -e ./orchid-api
ORCHID_CONFIG=examples/education/orchid.yml uvicorn orchid_api.main:app --port 8000

Or via the CLI:

pip install -e ./orchid -e ./orchid-cli
orchid chat interactive --config examples/education/orchid.yml

Install extra dependencies for PDF and PPTX export:

pip install reportlab python-pptx

Configuration walkthrough

orchid.yml sets the runtime — LLM, Qdrant for RAG, file uploads, and identity resolution:

# orchid.yml (trimmed)
agents:
config_path: examples/education/agents.yaml

llm:
#model: ollama/llama3.2
#ollama_api_base: http://host.docker.internal:11434
model: gemini/gemini-flash-latest
gemini_api_key: ${GEMINI_API_KEY}

auth:
dev_bypass: true
identity_resolver_class: examples.education.identity.EducationIdentityResolver

rag:
vector_backend: qdrant
qdrant_url: http://qdrant:6333
#embedding_model: ollama/nomic-embed-text
embedding_model: gemini/gemini-embedding-001
gemini_api_key: ${GEMINI_API_KEY}

upload:
#vision_model: ollama/minicpm-v
namespace: education-uploads
max_size_mb: 20
chunk_size: 1000
chunk_overlap: 200

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

# ...truncated

Agent configs define the four agents, their tools, three cross-agent skills, and the event pipeline:

# agents.yaml (trimmed)
version: "1"

defaults:
llm:
  model: "gemini/gemini-flash-latest"
  temperature: 0.2
rag:
  enabled: true
  k: 8
  max_context_chars: 4000

supervisor:
assistant_name: "Education Studio"

tools:
extract_concepts:
  class: "examples.education.tools.content.extract_concepts.ExtractConceptsTool"
  description: "Extract key concepts and headings from source material"
  parallel_safe: true
generate_questions:
  class: "examples.education.tools.content.generate_questions.GenerateQuestionsTool"
  description: "Generate quiz question scaffolds from extracted concepts"
  parallel_safe: true
validate_questions:
  class: "examples.education.tools.content.validate.ValidateQuestionsTool"
  description: "Validate questions for structure and duplication"
  parallel_safe: true
build_lesson_structure:
  class: "examples.education.tools.content.build_lesson.BuildLessonStructureTool"
  description: "Build a timed lesson-plan scaffold"
  parallel_safe: true
define_learning_objectives:
  class: "examples.education.tools.content.format_lesson.DefineLearningObjectivesTool"
  description: "Create Bloom-aligned learning objectives"
  parallel_safe: true
format_lesson_section:
  class: "examples.education.tools.content.format_lesson.FormatLessonSectionTool"
  description: "Format one lesson section as Markdown"
  parallel_safe: true
format_multiple_choice:
  class: "examples.education.tools.content.format_quiz.FormatMultipleChoiceTool"
  description: "Format multiple-choice questions as Markdown"
  parallel_safe: true
format_true_false:
# ... additional format and export tools

skills:
generate_quiz:
  description: "Analyze material, extract concepts, turn them into a quiz."
  steps:
    - agent: content-analyzer
      instruction: "Extract key concepts from the source material."
    - agent: quiz-generator
      instruction: "Generate a quiz grounded in those concepts."
    - agent: format-exporter
      instruction: "Format the quiz for the requested file format."
generate_lesson:
  description: "Analyze material and assemble a lesson plan."
  steps:
    - agent: content-analyzer
      instruction: "Identify concepts for a coherent lesson."
    - agent: lesson-builder
      instruction: "Build objectives, sections, and assessment."
    - agent: format-exporter
      instruction: "Format the lesson for the requested output format."
generate_full_package:
  description: "Produce quiz, lesson plan, and presentation exports."
  steps:
    - agent: content-analyzer
    - agent: quiz-generator
    - agent: lesson-builder
    - agent: format-exporter

agents:
content-analyzer:
  description: "Extracts headings and core concepts from uploaded material."
  tools: [extract_concepts]
  rag:
    namespace: education-source
  execution_hints:
    parallel_safe: true

quiz-generator:
  description: "Converts concepts into answerable questions."
  tools: [generate_questions, validate_questions]
  mini_agent:
    enabled: true
    max_count: 4
  rag:
    namespace: education
  execution_hints:
    parallel_safe: true

lesson-builder:
  description: "Turns concepts into a paced lesson structure."
  tools: [build_lesson_structure, define_learning_objectives, format_lesson_section]
  rag:
    namespace: education
  execution_hints:
    parallel_safe: true

format-exporter:
  description: "Exports content as PDF, DOCX, PPTX, Markdown, or text."
  tools: [format_multiple_choice, format_true_false, format_fill_blank, format_matching, generate_pdf, generate_docx, generate_pptx, generate_markdown, generate_txt]
  rag:
    enabled: false
  execution_hints:
    parallel_safe: false

events:
enabled: true
store:
  class: orchid_ai.events.backends.sqlite.SQLiteEventStorage
scheduler:
  class: orchid_ai.events.schedulers.apscheduler.APSchedulerBackend

schedules:
  - id: weekly-quiz-cron
    cron: "0 8 * * 1"
    trigger_id: weekly-quiz
    identity: { mode: service_account, name: quiz-bot }

triggers:
  - id: weekly-quiz
    "on": { signal: cron, cron: "0 8 * * 1" }
    emits:
      agent: quiz-generator
      prompt_template: |
        Generate a 10-question weekly review quiz.
      identity: { mode: service_account, name: quiz-bot }
      visibility: tenant
    parallelism: unbounded

  - id: chat-bound-generation
    "on":
      signal: education.generate
    emits:
      agent: content-analyzer
      prompt_template: |
        Extract key concepts from this source text.
        Source: {{payload.source_text}}
      identity: { mode: act_as_user, user_id_from: signal.user_id }
      respect_chat_binding: true
    parallelism: per_user

# ...truncated

What to look for

  • tools.<name>.class: "examples.education.tools.content.extract_concepts.ExtractConceptsTool" → tools are OrchidTool subclasses with declarative parameters_schema and parallel_safe flags; the framework resolves them by dotted import path at startup.
  • tools.*.parallel_safe → content-analysis tools are parallel_safe: true (stateless reads), while export tools that write files are parallel_safe: false (sequential I/O).
  • mini_agent.enabled: true on quiz-generator → multi-section source material fans out into parallel mini-agents, each drafting questions from one section. Custom decomposer_prompt and aggregator_prompt control the split/merge logic.
  • Three cross-agent skills (generate_quiz, generate_lesson, generate_full_package) → chain agents sequentially; each step passes context forward. The generate_full_package skill demonstrates a four-agent pipeline.
  • rag.namespace: education-source / education / education-exports → three distinct RAG namespaces isolate source content from generated quizzes from export artifacts.
  • events.schedules[].cron: "0 8 * * 1" → the APScheduler backend fires a weekly cron that triggers quiz-generator to produce a tenant-wide review quiz every Monday at 08:00.
  • identity.mode: service_account on the weekly quiz → the Bloom runs as quiz-bot, visible to all users in the tenant via visibility: tenant.
  • respect_chat_binding: true on chat-bound-generation → when a Bloom is triggered from within a chat context, its output is attached to that conversation rather than creating a standalone artifact.
  • format-exporter agent has rag.enabled: false → export formatting doesn't need retrieval; saves embedding and context overhead.
  • eywords → EducationIdentityResolver implements resolve_service_account() and mint_for_user(), enabling both cron-driven service-account flows and per-user event scoping.
  • upload: section in orchid.yml → file uploads target the education-uploads Qdrant namespace; the framework auto-chunks and indexes uploaded material at 1000-char chunks with 200-char overlap.

Related concepts