Enterprise AI Platform

An on-premise, containerized Enterprise AI Platform with multimodal input processing, a Meeting Intelligence module, agentic orchestration, RAG pipelines, and enterprise-grade observability. No external API calls at runtime — all models and services run inside an isolated Docker network.

Architecture reference: Project_Files/High Level design Diagram.png

---

Capabilities

Data Export Utility

This repo includes a migration/export CLI in export_tool/ for exporting PostgreSQL, Neo4j, Milvus, and Elasticsearch data into portable filesystem artifacts with manifests and SHA256 checksums.

cd export_tool
python main.py --output-dir ./exports

See export_tool/README.md for setup, filtering, compression, incremental export, and restore notes.

The matching import CLI lives in import_tool/ and restores those artifacts into target PostgreSQL, Neo4j, Milvus, and Elasticsearch environments:

cd import_tool
python main.py --input-dir ./exports --dry-run

---

Setup Paths

Jump to: Developer Laptop Setup · Full Stack Setup

---

Prerequisites

Ports that must be free before starting:

---

Quick Start

1. Prepare the environment file

cp .env.example .env

Open .env and replace every change_me_* placeholder with real values:

POSTGRES_PASSWORD=<strong-password>
REDIS_PASSWORD=<strong-password>
MINIO_ROOT_USER=admin
MINIO_ROOT_PASSWORD=<strong-password>
GRAFANA_ADMIN_PASSWORD=<strong-password>
SECRET_KEY=<random-32-char-string>

Generate a strong SECRET_KEY:

python3 -c "import secrets; print(secrets.token_hex(32))"

2. Build and start the stack

docker compose up --build -d

First run pulls all images and builds the application image — allow several minutes. Subsequent starts are fast.

docker compose logs -f api   # watch startup

3. Pull LLM models into Ollama

# Document RAG (light model)
docker exec enterprise_ai_ollama ollama pull qwen2.5:3b

# Meeting Intelligence (better quality, more RAM)
docker exec enterprise_ai_ollama ollama pull llama3.1:8b

4. Apply the database schema

The PostgreSQL init scripts in infra/docker/postgres/ run automatically on first container start. For an existing database, apply any new DDL files manually:

docker exec -i enterprise_ai_postgres psql -U aiuser -d enterprise_ai \
  < infra/docker/postgres/05_meeting_intelligence_ddl.sql

5. Upload documents and run the pipeline

# Upload PDFs from data/ to MinIO
make upload-data

# Extract → embed → index documents
make run-pipeline

---

Verifying the Stack

# API health
curl http://localhost:8000/health
# Expected: {"status":"ok"}

# Interactive API docs
open http://localhost:8000/docs

# Milvus health
curl http://localhost:9091/healthz

# PostgreSQL
docker exec enterprise_ai_postgres pg_isready -U aiuser -d enterprise_ai

# Redis
docker exec enterprise_ai_redis redis-cli -a <REDIS_PASSWORD> ping

# Ollama — list available models
curl http://localhost:11434/api/tags

# All containers at once
docker compose ps

---

Service Reference

---

Environment Variables

All variables are set in .env (copied from .env.example). Key variables:

See .env.example for the full list including development overrides.

---

Developer Laptop Setup

A slim Docker Compose override (docker-compose.dev.yaml) is provided for laptops. It:

• Uses a lighter image (~1 GB) — PyTorch, speechbrain, and heavy ML libs excluded
• Caps memory per service so the full stack fits in ~3–4 GB RAM
• Mounts src/ read-write for live reload (no rebuild on code changes)
• Makes Ollama and the monitoring stack opt-in via Docker Compose profiles

Prerequisites (laptop)

1. Environment file

cp .env.example .env

The example file includes inline dev/prod guidance. For a developer laptop, use the "Dev:" values noted in the comments (weak passwords, CPU models, lighter LLM).

2. Start the dev stack

make dev-up

First build takes a few minutes. Subsequent starts are fast.

3. Default services (dev)

Total: ~2.4 GB RAM cap across all containers.

4. Opt-in profiles

Tip: For Meeting Intelligence on a laptop, set dev-friendly model overrides in .env:

WHISPER_MODEL=whisper-base
WHISPER_DEVICE=cpu
WHISPER_COMPUTE=int8
MEETING_LLM_MODEL=qwen2.5:3b
MEETING_EMBEDDING_MODEL=all-MiniLM-L6-v2

5. Live reload

Edit any Python file under src/ — the API reloads automatically within 1–2 seconds.

make dev-logs   # watch reloads

Stopping

make dev-down   # stops everything including monitoring-profile containers

---

Meeting Intelligence

Upload any meeting recording and the platform processes it end-to-end in the background.

Supported formats

MP4, MKV, MOV, MP3, WAV, M4A, WEBM

Pipeline (8 stages)

1. Audio extraction — FFmpeg extracts 16kHz mono WAV from recording
2. Transcription — faster-whisper with VAD and hallucination filtering
3. Speaker diarization — pyannote assigns speaker labels per time segment
4. Merge — Whisper transcript segments are assigned speaker labels
5. Chunking — Sliding-window chunking with speaker-prefixed text
6. Vector indexing — BAAI/bge-large-en-v1.5 embeddings → Milvus (meeting_transcript_chunks, 1024-dim)
7. Text indexing — Chunks indexed in Elasticsearch (meeting_chunks)
8. LLM analysis — Summary (brief + key topics + decisions + open questions) and action items extracted

API endpoints

Offline / air-gapped deployment

pyannote models can be pre-downloaded and served from a local directory:

# Download once on an internet-connected machine
huggingface-cli download pyannote/speaker-diarization-3.1 \
    --local-dir /models/speaker-diarization-3.1 --token hf_...
huggingface-cli download pyannote/segmentation-3.0 \
    --local-dir /models/segmentation-3.0 --token hf_...

Then update config.yaml inside the pipeline dir to reference local model paths, and set:

PYANNOTE_MODEL=/models/speaker-diarization-3.1

No HuggingFace token is needed at runtime.

---

Document Pipeline Flow

data/ (PDFs, DOCX, XLSX)
  → MinIO (enterprise-ai-docs bucket)
  → Text extraction (pypdf / python-docx / openpyxl)
  → OCR on embedded images (tesseract ara+eng)
  → Embeddings (paraphrase-multilingual-MiniLM-L12-v2, 384-dim)
  → Milvus  — vector embeddings for semantic search  (document_chunks)
  → Elasticsearch — keywords + BM25 full-text search (document_chunks)
  → Postgres — document registry, pipeline run status, audit events

---

Project Structure

Enterprise_AI/
├── CLAUDE.md                  # Architecture and developer guide
├── Dockerfile                 # Multi-stage Python 3.11 build (dev + prod paths)
├── docker-compose.yaml        # Full on-premise service stack
├── docker-compose.dev.yaml    # Developer laptop override (slim, memory-capped)
├── Makefile                   # Convenience targets
├── requirements.txt           # Full Python dependencies (includes PyTorch, ML libs)
├── requirements.dev.txt       # Slim dependencies for dev image (no PyTorch/CV)
├── .env.example               # Environment variable template (includes dev/prod guidance inline)
├── src/
│   ├── api/                   # FastAPI application (docs, meetings, query, pipeline routes)
│   ├── agents/                # Agent orchestration
│   ├── multimodal/            # Audio / video / document processing
│   ├── services/              # LLM, RAG, embeddings, Meeting Intelligence pipeline
│   ├── workers/               # Celery task workers (meeting_pipeline.py)
│   ├── data/                  # Milvus, Elasticsearch, PostgreSQL, Redis adapters
│   └── observability/         # Logging, tracing, metrics
├── infra/docker/              # Service config files (OTel, Prometheus, Grafana, Tempo)
├── scripts/                   # Data management + living documentation pipeline
├── templates/                 # Jinja2 templates for doc generation
└── tests/                     # Unit, integration, e2e

---

Frontend (React + Vite)

The frontend/ directory contains a React 18 + TypeScript + Tailwind UI.

Local development (recommended):

make frontend-install   # first time
make frontend           # http://localhost:3002

Production (Docker, port 3010):

docker compose --profile frontend up -d frontend

Features

• Streaming responses via SSE
• Arabic RTL auto-detection
• Source citation cards with document name, page, relevance score
• Document browser sidebar
• Top-K slider for retrieved context chunks

---

Stopping and Resetting

# Stop all services (preserves data volumes)
make full-down

# Stop and wipe all data volumes (full reset)
docker compose down -v

Warning: -v permanently deletes all stored data including the vector store, database, and uploaded files.

---

Troubleshooting

Milvus takes a long time to become healthy Allow up to 60 seconds on first start. If it never becomes healthy:

docker compose logs milvus
docker compose logs etcd

Port already in use

# Linux/macOS
lsof -i :<PORT>
# Windows
netstat -ano | findstr :<PORT>

Meeting processing stuck at "processing"

make dev-logs   # check worker output
docker logs enterprise_ai_worker

api container exits immediately The API depends on PostgreSQL, Redis, Milvus, and Elasticsearch being healthy:

docker compose ps
docker compose logs postgres redis milvus elasticsearch

MinIO bucket not found

docker exec enterprise_ai_minio mc alias set local http://localhost:9000 $MINIO_ROOT_USER $MINIO_ROOT_PASSWORD
docker exec enterprise_ai_minio mc mb local/enterprise-ai-docs
docker exec enterprise_ai_minio mc mb local/enterprise-ai-meetings

Ollama model not found

docker exec enterprise_ai_ollama ollama pull qwen2.5:3b
docker exec enterprise_ai_ollama ollama pull llama3.1:8b