Gallery Curator

A multi-session art gallery curator that remembers artists, exhibitions, and decisions across sessions using Orchid's full conversation summarization pipeline.

What this demonstrates

The Gallery Curator is a single GenericAgent that helps gallery directors plan exhibitions, research artists, track auction results, and record curatorial decisions. It's purpose-built to showcase every feature in Orchid's conversation summarization pipeline (Phases 1–5).

Feature	Where configured	What happens
Incremental running summary	`supervisor.memory.strategy: running_summary`	Each new turn extends the summary instead of re-summarizing from scratch
Structured JSON entity extraction	`supervisor.memory.structured_output: true`	Artists, venues, exhibitions tracked as typed entities; deduplicated across turns
RAG-augmented semantic retrieval	`supervisor.memory.strategy: rag_augmented`	Past turns retrieved from Qdrant via embedding similarity to the current query
Configurable compression prompts	`agents[].prompt_sections`	Gallery-specific instructions for the compression LLM
Smart message truncation	`supervisor.memory.truncation_strategy: middle`	Long auction listings truncated keeping start (which item?) and end (how much?)
SQLite persistence	`storage.class` + `persist_summary: true`	Summaries survive restarts via `conversation_summaries` table
Graceful degradation	Qdrant unavailable → NullVectorReader	`rag_augmented` silently degrades to `running_summary`-only

Run it

# From the monorepo root
pip install -e ./orchid -e ./orchid-api
ORCHID_CONFIG=examples/gallery-curator/orchid.yml \
  uvicorn orchid_api.main:app --port 8000

Or validate the YAML:

pip install -e ./orchid -e ./orchid-cli
orchid config validate examples/gallery-curator/agents.yaml

Or via the CLI (no API server needed):

orchid chat interactive --config examples/gallery-curator/orchid.yml

Configuration walkthrough

The key block of agents.yaml — the supervisor.memory section:

supervisor:
  history_summary_enabled: true
  history_max_turns: 30
  history_max_chars: 1200

  memory:
    strategy: "running_summary"
    summary_recent_turns: 8
    summary_model: ollama/llama3.2
    structured_output: true
    persist_summary: true
    # -- RAG memory (active when strategy=rag_augmented) --
    rag_k: 5
    rag_similarity_threshold: 0.5
    store_turns: true
    # -- Truncation --
    truncation_strategy: "middle"
    truncation_max_chars: 800

Per-agent compression prompt overrides in prompt_sections:

agents:
  gallery-curator:
    prompt_sections:
      summary_compression_system_prompt: >
        You are a gallery curator conversation summarizer. Extract structured
        information from exhibition planning discussions...

      summary_compression_user_prompt: >
        Summarise the following gallery planning conversation in structured JSON...
        {transcript}

      summary_extension_user_prompt: >
        Given the existing summary below and the new gallery planning messages...
        Existing summary: {existing_summary}
        New messages: {new_messages}

Strategy comparison

Strategy	Summary persistent?	Uses RAG?	Best for
`none`	No	No	Short-lived chats, backward compat
`running_summary`	Yes (SQLite/Postgres)	No	Long sessions, O(n) cost
`rag_augmented`	Yes	Yes	Multi-session with semantic recall

Switch to rag_augmented in one line:

memory:
  strategy: "rag_augmented"  # was "running_summary"

Everything else stays the same — the OrchidRAGConversationMemory class automatically adds Qdrant retrieval on top of the running summary.

Simulated conversation flow

Turn 1:  "I'm planning the spring exhibition..."
         → No prior context. First summary created.

Turn 5:  "What was the budget for Ruth Asawa's shipping?"
         → Running summary covers turns 1–4.
         → With rag_augmented: Qdrant retrieves past "budget" turns.

Turn 15: "Remind me of all Venice decisions."
         → Summary has been incrementally extended 10+ times (O(n) cost).
         → Structured JSON tracks 8+ entities with deduplication.
         → With rag_augmented: top-5 Venice turns retrieved from Qdrant.
         → LLM receives: [RAG turns] + [incremental summary] + [last 8 verbatim].

Files

examples/gallery-curator/
├── README.md       # Full documentation with entity examples
├── orchid.yml      # Infrastructure: SQLite, Qdrant, Ollama
└── agents.yaml     # Agent + supervisor + memory config

No custom Python agent classes — everything is GenericAgent driven by YAML.