Moncey Concierge

Live at: https://concierge.aws.monce.ai
Chat UI: https://concierge.aws.monce.ai/ui

Internal memory and intelligence layer for Monce AI. Tracks extraction pipeline activity, answers analytical questions from pre-computed digests, and bridges to Snake for synonym management.

Architecture

Route53 → EC2 (nginx/SSL → gunicorn) → FastAPI + Bedrock Sonnet
                                              │
                         ┌────────────────────┼──────────────────┐
                    Memory (JSON)        monce_db (S3)     snake.aws.monce.ai
                    - memories.json      - extractions      - article synonyms
                    - conversations.json - stats            - client synonyms
                    - digests.json                          - rebuild triggers

Quick Start

# Chat
curl -X POST https://concierge.aws.monce.ai/chat \
  -H 'Content-Type: application/json' \
  -d '{"message": "What are the top clients this week?"}'

# Ingest last 14 days of extractions
curl -X POST https://concierge.aws.monce.ai/ingest \
  -H 'Content-Type: application/json' \
  -d '{"days": 14}'

# Search memories
curl 'https://concierge.aws.monce.ai/search?q=SGD'

# Add article synonym to Snake
curl -X POST https://concierge.aws.monce.ai/snake/synonym \
  -H 'Content-Type: application/json' \
  -d '{"text": "6mm", "num_article": "1006", "factory_id": "3"}'

# Add client synonym to Snake
curl -X POST https://concierge.aws.monce.ai/snake/synonym_client \
  -H 'Content-Type: application/json' \
  -d '{"text": "DUBOS MATERIAUX", "numero_client": "565", "factory_id": "4"}'

API Endpoints

How Concierge Answers Questions

Concierge uses a 3-layer context system so Sonnet can answer precisely:

1. Digests — Pre-computed aggregates from ALL extraction data (top clients, daily volumes, glass types, matching quality, weekly rankings). Always included. Compact.
2. Search results — Memories keyword-matched to the user's question (up to 20). Targeted.
3. Recent memories — Last 10 raw ingestions. For "what just happened" questions.

When you call /ingest, digests are auto-recomputed. You can also manually trigger /digest POST.

Feeding Data to Concierge

From monce_db (extractions)

# Ingest last 14 days, all factories
curl -X POST https://concierge.aws.monce.ai/ingest \
  -d '{"days": 14}'

# Ingest specific factory
curl -X POST https://concierge.aws.monce.ai/ingest \
  -d '{"days": 14, "factory": "VIP"}'

# Only verified extractions
curl -X POST https://concierge.aws.monce.ai/ingest \
  -d '{"days": 14, "status": "verified"}'

Deduplicates by extraction ID — safe to call repeatedly.

Manual memories

# Remember something
curl -X POST https://concierge.aws.monce.ai/remember \
  -d '{"text": "Factory 4 had a major outage today", "tags": ["incident", "VIP"]}'

Programmatic best practices

To make Concierge the effective memory of Monce AI:

1. Tag everything. Tags enable filtering and weighted search. Use consistent tags: extraction, synonym, incident, factory names.

2. Ingest regularly. Set up a cron or call /ingest daily. Concierge deduplicates, so overcalling is fine.

3. Use /remember for non-extraction events. Deployments, incidents, configuration changes — anything Sonnet should know about when answering questions.

4. Let digests do the heavy lifting. Don't ask Concierge to count raw memories — digests pre-compute totals, rankings, and trends. If you need a new aggregate, add it to compute_digests() in memory.py.

5. Search before asking. For programmatic lookups, use /search?q=keyword instead of /chat. It's faster and doesn't consume Bedrock tokens.

Snake Synonym Integration

Concierge can push synonyms directly to snake.aws.monce.ai (article matching service). This is useful when extraction analysis reveals missing or incorrect synonym mappings.

Article synonyms

curl -X POST https://concierge.aws.monce.ai/snake/synonym \
  -d '{"text": "PLANILUX 4MM", "num_article": "1004", "factory_id": "3"}'

Client synonyms

curl -X POST https://concierge.aws.monce.ai/snake/synonym_client \
  -d '{"text": "SAINT GOBAIN PARIS", "numero_client": "7890", "factory_id": "4"}'

Batch workflow

# Push multiple synonyms without rebuilding each time
curl -X POST https://concierge.aws.monce.ai/snake/synonyms_batch \
  -d '{
    "synonym_type": "article",
    "synonyms": [
      {"text": "6mm", "num_article": "1006", "factory_id": "3"},
      {"text": "FLOAT 6", "num_article": "1006", "factory_id": "3"},
      {"text": "8mm clair", "num_article": "1008", "factory_id": "4"}
    ]
  }'

Batch adds all synonyms with trigger_rebuild=false, then calls /rebuild_all once at the end.

Every synonym action is logged as a Concierge memory with tags [synonym, article/client, factory_id].

Claude Code Sync

To work on Concierge with Claude Code:

git clone git@github.com:Monce-AI/concierge.aws.monce.ai.git
cd concierge.aws.monce.ai

File structure

concierge.aws.monce.ai/
  api/
    __init__.py
    main.py          # FastAPI app entry
    config.py        # Env var config (Bedrock, data dir)
    routes.py        # All endpoints
    sonnet.py        # Bedrock Sonnet caller + system prompt
    memory.py        # Memory CRUD + digest engine + search
    ingest.py        # monce_db ingestion
    snake.py         # Snake API client (synonyms + rebuild)
    static/
      index.html     # Landing page
      ui.html        # Chat interface
  terraform/
    main.tf          # EC2 + SG + Route53
    deploy.sh        # Rsync + systemd + nginx
  setup.py

Deploy

cd terraform
./deploy.sh          # or ./deploy.sh <ip>

Environment variables (on server at /opt/concierge/.env)

AWS_BEARER_TOKEN_BEDROCK=...   # Bedrock access
MONCE_S3_ACCESS_KEY=...        # monce_db S3 access
MONCE_S3_SECRET_KEY=...        # monce_db S3 secret

Adding new capabilities

1. New data source: Add an ingestion function in ingest.py, add a route in routes.py
2. New digest type: Add computation logic in memory.py → compute_digests()
3. New external service: Create a module (like snake.py), add routes
4. Changing Sonnet's behavior: Edit SYSTEM_PROMPT in sonnet.py

Infrastructure