memorai 0.1.0

Conversation memory as a knowledge graph — pipeline, retrieval, and generation

MemOrai

Build knowledge graphs from conversations and answer questions using retrieval-augmented generation over a knowledge graph.

---

📦 Installation

From PyPI

pip install memorai

With FastAPI backend extras

pip install "memorai[backend]"

From source (editable install)

git clone https://github.com/memorai/memorai.git
cd memorai
pip install -e .
# or with backend extras:
pip install -e ".[backend]"

---

⚙️ Configuration

Create a .env file (or set environment variables) before running:

# LLM Configuration (OpenAI-compatible endpoint)
LLM_API_KEY=your-api-key-here
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL=google/gemini-2.0-flash-001

# Embedding Model (HuggingFace model name)
EMBEDDING_MODEL=BAAI/bge-m3

# Database Configuration (Neo4j)
# Use Aura DB or a local Neo4j instance
NEO4J_URI=neo4j+s://your-id.databases.neo4j.io
NEO4J_USER=neo4j
NEO4J_PASSWORD=your-password

---

🚀 Quick Start

MemOrai uses Neo4j as a unified storage backend. It acts both as a Document Store (for pipeline states) and a Knowledge Graph. No local files are needed for retrieval once indexed.

Python API

import memorai

# 1. Initialize a conversation scope
memorai.create_conversation(
    conversation_id="alice-bot",
    name="Alice - Support Bot",
)

# 2. Index conversation history (builds graph in Neo4j)
history = [
    {"role": "user", "content": "My name is Alice and I live in Hanoi."},
    {"role": "assistant", "content": "Nice to meet you, Alice!"},
]

memorai.index(
    history=history,
    conversation_id="alice-bot",
    session_id="alice-session-001",
    update=True,
    fast_mode=True,
)

# 3. Retrieve using Graph Vector Search
result = memorai.retrieve(
    query="Where does Alice live?",
    conversation_id="alice-bot",
)
print(result["top_turn_contents"])

Notes:

• conversation_id isolates tenant data in Neo4j.
• session_id lets you append incremental chat batches inside one conversation scope.
• fast_mode=True runs low-latency indexing (skips heavy post-processing).

CLI — Full pipeline

# Run full pipeline from a JSON file
memorai pipeline \
    --input_json data/conversations.json \
    --output_dir output \
    --save_embeddings \
    --cleanup

# Answer a single question
memorai qa \
    --data_path output/graph_db/session-001 \
    --query "Where does Alice live?"

# Batch QA
memorai qa-batch \
    --questions_file questions.json \
    --data_path output/graph_db/session-001 \
    --output answers.json

---

📋 CLI Commands

Pipeline commands

Command	Description
memorai segment	Segment conversations into turns
memorai filter	Filter important messages
memorai triplets	Extract knowledge triplets
memorai entities	Generate entity descriptions
memorai summarize	Summarize segments
memorai graph	Build knowledge graph
memorai pipeline	Run full pipeline end-to-end

Post-processing commands

Command	Description
memorai segment-chunk-map	Export segment → chunk mapping
memorai consolidate-turns	Deduplicate turn IDs
memorai rebuild-graph	Rebuild graph after consolidation
memorai embed-turns	Add turn embeddings
memorai embed-entities	Add entity embeddings
memorai embed-triplets	Add triplet embeddings
memorai embed-summaries	Add summary embeddings

QA commands

Command	Description
memorai retrieve	Retrieve relevant nodes from KG
memorai qa	Answer a single question
memorai qa-batch	Answer a batch of questions

---

🏗️ Architecture

Conversation History
        │
        ▼
  ┌─────────────┐
  │  Segmenter  │  Split into semantic turns
  └──────┬──────┘
         │
  ┌──────▼──────┐
  │   Filter    │  Remove low-signal messages
  └──────┬──────┘
         │
  ┌──────▼──────────┐
  │ TripletExtractor│  Extract (entity, relation, entity)
  └──────┬──────────┘
         │
  ┌──────▼──────────┐
  │ EntityDescriptor│  Describe entities in context
  └──────┬──────────┘
         │
  ┌──────▼──────┐
  │ GraphBuilder│  Build Knowledge Graph in Neo4j
  └──────┬──────┘
         │
  ┌──────▼────────┐
  │ Neo4jRetriever│  Vector Search + Cypher Traversal
  └──────┬────────┘
         │
  ┌──────▼────────┐
  │ AnswerGenerator│  RAG over Neo4j context
  └───────────────┘

---

🔧 Development

# Install with dev extras
pip install -e ".[dev]"

# Run tests
pytest

# Build distribution
make build

# Check package
twine check dist/*

# Publish to PyPI
make publish

---

📤 Publish Guide (DIY)

Use this section when you want to publish manually.

1. Prepare account + API tokens

1. Create an account on PyPI and TestPyPI.
2. Create API token on TestPyPI (for dry-run upload).
3. Create API token on PyPI (real release).
4. Keep tokens in env vars (recommended):

export TWINE_USERNAME=__token__
export TWINE_PASSWORD=pypi-<your-token>

2. Build package artifacts

python -m pip install --upgrade build twine
rm -rf dist build *.egg-info
python -m build
python -m twine check dist/*

Expected artifacts:

• dist/memorai-<version>.tar.gz
• dist/memorai-<version>-py3-none-any.whl

3. Upload to TestPyPI first

export TWINE_USERNAME=__token__
export TWINE_PASSWORD=pypi-<your-testpypi-token>
python -m twine upload --repository testpypi dist/*

Install test package:

python -m pip install \
    --index-url https://test.pypi.org/simple/ \
    --extra-index-url https://pypi.org/simple \
    memorai

4. Publish to real PyPI

export TWINE_USERNAME=__token__
export TWINE_PASSWORD=pypi-<your-pypi-token>
python -m twine upload dist/*

5. Verify release

python -m pip install --upgrade memorai
python -c "import memorai; print(memorai.__version__)"

Common issues

• File already exists: bump version in pyproject.toml and memorai/__init__.py, then rebuild.
• 403 invalid token: ensure token scope matches target index (PyPI vs TestPyPI).
• Long README render errors: run python -m twine check dist/* before upload.

---

📄 License

MIT — see LICENSE.