Tether - Source-Grounded Workplace Memory

Tether is a multi-modal Retrieval-Augmented Generation (RAG) system for searching workplace meetings and internal documents with precise source attribution.

Overview

Tether turns meeting recordings and documents into a searchable organizational memory layer. Teams can ask who said what, recover decisions, find action items, and verify facts against timestamped meeting transcripts and document citations.

Key Features

• Multi-Modal Search: Unified search across meeting audio and internal documents
• Hybrid Retrieval: Combines keyword search (BM25) with semantic vector search for accurate recall
• Speaker-Aware Evidence: Preserves speaker labels and meeting timestamps for "who said what" queries
• Precise Citations: Every answer includes source attribution, including document pages and audio timestamps
• Quality Evaluation: Built-in RAGAS metrics for faithfulness and relevance scoring

Technology Stack

• LLM: Groq API
• Embeddings: Sentence Transformers (all-MiniLM-L6-v2, local)
• Vector Database: Qdrant
• Transcription: OpenAI Whisper (local)
• Speaker Diarization: pyannote.audio
• Web Interface: Streamlit

---

Installation

Prerequisites

• Python 3.10 or higher
• Docker (for Qdrant)
• 8GB RAM minimum (16GB recommended)
• Groq API key

Setup Steps

1. Clone Repository

git clone <repository-url>
cd tether

2. Create Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Dependencies

With uv:

uv sync --extra dev

Or with pip:

pip install --upgrade pip
pip install -e ".[dev]"

4. Start Qdrant Vector Database

docker run -d -p 6333:6333 -v $(pwd)/qdrant_storage:/qdrant/storage qdrant/qdrant

Verify Qdrant is running:

curl http://localhost:6333/health

5. Configure Environment

cp .env.example .env

Edit .env and add your Groq API key:

GROQ_API_KEY=your_groq_api_key_here

Optional: Add HuggingFace token for speaker diarization:

ENABLE_DIARIZATION=true
HF_TOKEN=your_huggingface_token_here

If you want the optional diarization dependency installed, use:

uv sync --extra dev --extra diarization

6. Prepare Workspace Data

Create the directory structure and add your files:

mkdir -p data/workspace/documents
mkdir -p data/workspace/meetings

Place your files:

• PDFs in data/workspace/documents/
• Meeting audio files (MP3, WAV, M4A) in data/workspace/meetings/

7. Ingest Workspace Content

python scripts/ingest.py

This will:

• Extract text from PDFs and chunk with metadata
• Transcribe meeting audio with Whisper while preserving timestamps
• Perform speaker diarization if enabled
• Generate embeddings and index content in Qdrant

Expected output:

============================================================
INGESTION SUMMARY
============================================================
PDFs processed:     1
Audio processed:    1
Total chunks:       156

All files processed successfully!
============================================================

---

Usage

Web Interface

Start the Streamlit app:

streamlit run streamlit_app/app.py

Navigate to http://localhost:8501 in your browser.

Command Line Interface

Single query:

python scripts/query.py --query "What did we decide about the launch timeline?"

With filters:

python scripts/query.py \
  --query "What did Alex say about onboarding?" \
  --response-mode who_said_what \
  --mode hybrid \
  --filter-source audio \
  --filter-speaker SPEAKER_00

Interactive mode:

python scripts/query.py --interactive

Run Evaluation

python scripts/evaluate.py

This generates a quality report with RAGAS metrics in outputs/evaluation_report.json.

Run Tests

pytest tests/ -v

---

Example Questions

• "Who suggested delaying the release?"
• "What did we decide about pricing?"
• "What action items came out of the onboarding meeting?"
• "Which document supports the API rate-limit decision?"
• "Summarize all discussions about customer migration risk."

Response Modes

Tether supports task-specific answer modes:

• answer: General source-grounded Q&A.
• who_said_what: Speaker-aware recall with timestamped citations.
• decisions: Decision extraction and rationale.
• action_items: Task, owner, and due-date extraction from indexed sources.
• fact_check: Support, contradiction, or insufficient-evidence checks against meetings and documents.

---

Project Structure

tether/
├── config/              # Configuration settings
├── src/
│   ├── ingestion/       # PDF and audio processing
│   ├── storage/         # Vector database wrapper
│   ├── retrieval/       # Hybrid search engine
│   ├── generation/      # RAG answer generation
│   ├── evaluation/      # Quality metrics
│   └── utils/           # Logging and error handling
├── scripts/             # CLI tools (ingest, query, evaluate)
├── streamlit_app/       # Web interface
├── tests/               # Unit tests
├── data/                # Workspace files (not in repo)
└── outputs/             # Query results and reports

---

Portfolio Positioning

Tether is not positioned as a generic meeting summarizer. It is an evidence-first workplace memory system:

Every workplace answer, tethered to its source.

The goal is to recover decisions, action items, and facts from meeting recordings and internal documents with transparent citations and measurable answer quality.