ava

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

(★ 20)

AI voice assistant that answers your phone calls with a human like real persona, holds natural multilingual conversations (GPT-4o + ElevenLabs), and keeps you in the loop via Signal with realtime midcall instructions.

ava Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# AVA – AI Voice Assistant

> **AVA** answers your calls when you can't, holds a natural conversation with a human-like persona, and keeps you in the loop via Signal. You can send live instructions mid-call from your phone.

---

## Architecture Overview

```mermaid
graph TB
 subgraph External["EXTERNAL SERVICES"]
 Twilio["Twilio Voice / PSTN STT (Gather) Record Webhooks"]
 OpenAI["OpenAI GPT-4o (conversation) Whisper async TTS (fallback)"]
 ElevenLabs["ElevenLabs TTS (primary voice) eleven_turbo_v2_5"]
 end

subgraph Docker["DOCKER HOST (your server)"]
 subgraph Ingress["INGRESS (choose one)"]
 Caddy["Caddy :443/:80 Let's Encrypt auto HTTPS"]
 Cloudflared["Cloudflare Tunnel outbound, no open ports"]
 end

subgraph AVA["AVA (FastAPI :8000)"]
 Main["main.py Call routing Twilio hooks Whisper async Rate limiter Audio serve"]
 Conv["conversation.py GPT-4o / Groq Streaming Meta parsing Summarizer"]
 TTS["tts.py ElevenLabs → OpenAI → Polly (fallback) Cache (MD5) Circuit breaker"]
 Owner["owner_channel.py Signal notify Signal poll (3s) Slash commands Instructions"]
 Contact["contact_lookup.py contacts.json Twilio CNAM E.164 normalize Lang from prefix"]
 I18n["i18n.py 11+ languages Signal templates Polly voices Twilio codes"]
 end

SignalCLI["signal-cli :8080 REST API Native mode Self-hosted"]

subgraph Volumes["Persistent Volumes"]
            TTSCache["tts_cache (MP3s)"]
            CallData["/data/calls/ (JSON)"]
            Contacts["/data/contacts.json"]
            SignalData["signal_data"]
        end
    end

OwnerPhone["Owner's Phone (Signal app)"]

Main <--> Conv
 Main <--> TTS
 Main <--> Owner
 Main <--> Contact
 Conv <--> I18n
 Main <--> I18n

Owner -->|"HTTP (ava-net)"| SignalCLI
 SignalCLI <-->|"Signal protocol"| OwnerPhone

TTS --> TTSCache
    Main --> CallData
    Contact --> Contacts
    SignalCLI --> SignalData

style External fill:#f9f0ff,stroke:#7c3aed
    style Docker fill:#f0f9ff,stroke:#2563eb
    style AVA fill:#ecfdf5,stroke:#059669
    style Ingress fill:#fef3c7,stroke:#d97706
    style Volumes fill:#fef2f2,stroke:#dc2626
```

---

## Call Flow (detailed sequence)

```mermaid
sequenceDiagram
    participant Caller as Caller's Phone
    participant Twilio as Twilio (PSTN + STT)
    participant AVA as AVA Server
    participant GPT as LLM (GPT-4o / Groq)
    participant TTS as ElevenLabs / OpenAI TTS
    participant Signal as Owner (Signal)

Caller->>Twilio: Dials owner (call forwarded)
 Twilio->>AVA: POST /twilio/incoming (CallSid, From, To)

Note over AVA: Contact lookup (local/CNAM) Detect lang from phone prefix (+41→de-CH, +48→pl-PL)

AVA-->>Signal: 📞 Incoming call notification

alt Contact has lang override
 AVA->>TTS: Generate greeting TTS
 AVA->>Twilio: TwiML: Gather + Play (known language)
 Twilio->>Caller: Plays greeting audio
 else Unknown contact
 AVA->>Twilio: TwiML: Record + Say "Which language do you prefer?"
 Twilio->>Caller: Plays question audio
 Caller->>Twilio: Speaks
 Twilio->>AVA: POST /twilio/first_response
 Note over AVA: OpenAI Whisper transcribes and detects actual language
 end

loop Max 10 exchanges
 Caller->>Twilio: Speaks
 Twilio->>AVA: POST /process_speech (SpeechResult, Confidence)

Note over AVA: Pop Signal instructions

opt Owner sent instruction
 Signal-->>AVA: "tell him I'll call back"
 Note over AVA: Inject [RELAY_TO_CALLER: ...] into GPT user message
 end

AVA->>GPT: Stream LLM (user text + instructions)
        GPT-->>AVA: Sentence chunks (streaming)

Note over AVA: TTS pipeline: start TTS on 1st sentence while GPT still generates the rest

AVA->>TTS: TTS sentence 1 (parallel)
        TTS-->>AVA: MP3 URL
        AVA->>TTS: TTS remaining sentences

Note over AVA: Parse meta JSON end_call, urgency, topic, caller_name, lang

opt GPT switched language
 Note over AVA: Update STT language for next Gather e.g. de-CH → pl-PL
 end

AVA->>Twilio: TwiML: Gather + Play (updated STT language)
 Twilio->>Caller: Plays response audio

opt Every 4 transcript entries
 AVA-->>Signal: 📞 Live update (topic, last 6 lines)
 end
 end

Note over AVA: end_call=true OR END_CALL_NOW from owner

AVA->>Twilio: TwiML: Play + Hangup
    Twilio->>Caller: Goodbye + disconnect

Twilio->>AVA: POST /twilio/status CallStatus=completed

AVA->>GPT: Summarize full transcript
    GPT-->>AVA: Summary text
    AVA-->>Signal: 📋 Call summary + priority
    AVA-->>Signal: 📝 Full transcript

Note over AVA: Save JSON to /data/calls/ Cleanup after 90s delay
```

---

## Timeouts & Limits

| Parameter | Value | Location | Description |
|-----------|-------|----------|-------------|
| `speech_timeout` | **1 s** | `main.py` (all Gather calls) | Silence after speech ends before Twilio fires callback |
| `enhanced` | `true` | `main.py` (Gather) | Use enhanced STT model for better accuracy |
| LLM `max_tokens` | **180** | `conversation.py` | Max response length per turn |
| GPT `temperature` | **0.75** | `conversation.py` | Creativity level for responses |
| Summary `max_tokens` | **400** | `conversation.py` | Max summary length |
| Summary `temperature` | **0.2** | `conversation.py` | Low creativity for factual summaries |
| Context window | **last 20 messages** | `conversation.py` | Sliding window of conversation history |
| Hard turn limit | **10 exchanges** | `conversation.py` | AVA wraps up after 10 user turns |
| Wrap-up warning | **8+ exchanges** | `conversation.py` | System prompt warns AVA to end soon |
| ElevenLabs timeout | **15 s** | `tts.py` (httpx) | HTTP timeout for TTS API |
| ElevenLabs circuit breaker | **10 min** | `tts.py` | Disable after 401/403/429, auto-reset |
| Signal poll interval | **3 s** | `main.py` / `owner_channel.py` | How often AVA checks for new Signal messages |
| Signal HTTP timeout | **10 s** | `owner_channel.py` (httpx) | Timeout for Signal API calls |
| CNAM lookup timeout | **5 s** | `contact_lookup.py` (httpx) | Twilio CNAM API timeout |
| Rate limiter | **30 req/min** per IP | `main.py` | Sliding window, auto-cleanup every 5 min |
| Rate limiter cleanup | **5 min** | `main.py` | Stale entry eviction interval |
| Call state cleanup | **90 s** after end | `main.py` | Delayed cleanup of in-memory call state |
| TTS cache | **no expiry** | `tts.py` | MD5(lang:text) keyed, persists in Docker volume |
| Seen Signal timestamps | **500 entries** | `owner_channel.py` | Deque for deduplication |

---

## Language Detection & Switching

```mermaid
flowchart TD
 Start([CALL START]) --> Prefix["Phone prefix detection +41 → de-CH +48 → pl-PL +44 → en-GB (52 prefixes)"]

Prefix --> ContactCheck{Contact has lang override?}
 ContactCheck -->|Yes| ContactLang["Use contact language contacts.json e.g. {name: ..., lang: pl}"]
 ContactCheck -->|No| Record["Twilio Record Ask language preference in prefix language"]

ContactLang --> Gather
 Record --> Whisper["OpenAI Whisper API Transcribes audio & detects actual language"]

Whisper --> GPT

Gather["Twilio STT Gather language = detected locale speech_timeout = 1s enhanced = true"]

Gather --> Speech["SpeechResult (text)"]
 Speech --> GPT["GPT-4o processes text Responds in caller's language Returns meta with lang: pl"]

GPT --> Switch{LLM lang ≠ current STT?}
 Switch -->|Yes| Update["Switch STT language for NEXT Gather e.g. de-CH → pl-PL"]
 Switch -->|No| Keep["Keep current STT language"]

Update --> Gather
    Keep --> Gather

style Start fill:#059669,color:#fff
    style Record fill:#2563eb,color:#fff
    style Whisper fill:#7c3aed,color:#fff
    style Gather fill:#2563eb,color:#fff
    style GPT fill:#7c3aed,color:#fff
    style Switch fill:#d97706,color:#fff
```

> **Important limitation**: Twilio STT only supports **one language per Gather**. If the caller speaks Polish but STT is set to German, the transcript will be garbled. The GPT model analyzes the garbled text and switches the language via the `meta` block for the **next** turn.

---

## TTS Provider Chain

```mermaid
flowchart TD
 Input["Text to speak"] --> Cache{Disk cache hit? key = MD5 lang:text}

Cache -->|Yes| Serve["Return cached URL PUBLIC_URL/audio/hash.mp3"]
 Cache -->|No| ELCheck{ElevenLabs available? API key set? Circuit breaker OK?}

ELCheck -->|Yes| EL["ElevenLabs API voice_id (env) model_id (env) timeout: 15s"]
 ELCheck -->|No| OpenAI

EL -->|Success| Save["Save to cache Return URL"]
 EL -->|Fail| OpenAI["OpenAI TTS model: tts-1 voice: OPENAI_TTS_VOICE (default: nova)"]

OpenAI -->|Success| Save
 OpenAI -->|Fail| Polly["Twilio Say (Polly) Last resort Built-in voice"]

EL -->|"401/403/429"| CB["Circuit Breaker Disable ElevenLabs for 10 minutes"]
 CB --> OpenAI

Save --> Done([Audio URL returned])
    Polly --> Done2([TwiML Say fallback])

style Input fill:#2563eb,color:#fff
    style EL fill:#7c3aed,color:#fff
    style OpenAI fill:#059669,color:#fff
    style Polly fill:#dc2626,color:#fff
    style CB fill:#d97706,color:#fff
    style Done fill:#059669,color:#fff
```

---

## Signal Communication Flow

```mermaid
sequenceDiagram
    participant Owner as Owner's Signal
    participant CLI as signal-cli REST API
    participant AVA as AVA Server

loop Every 3 seconds
        AVA->>CLI: GET /v1/receive
        CLI-->>AVA: [] (no messages)
    end

Note over AVA: INCOMING CALL

AVA->>CLI: POST /v2/send
 CLI->>Owner: 📞 Incoming call From: Jan (+48...) 🌐 pl-PL

Owner->>CLI: "tell him I'll call back"
 AVA->>CLI: GET /v1/receive
 CLI-->>AVA: [message data]
 Note over AVA: Queue instruction for active call

AVA->>CLI: POST /v2/send
    CLI->>Owner: ✅ AVA will tell the caller

Note over AVA: Next speech turn: inject instruction into GPT context

Note over AVA: After 4 transcript entries

AVA->>CLI: POST /v2/send
 CLI->>Owner: 📞 Call in progress 🟡 Topic: invoice dispute Last 6 lines of transcript

Note over AVA: CALL ENDS

AVA->>CLI: POST /v2/send
 CLI->>Owner: 📋 Call summary Priority + AI summary

AVA->>CLI: POST /v2/send
    CLI->>Owner: 📝 Full transcript
```

### Slash commands (no active call needed)

| Command | Description |
|---------|-------------|
| `/ping` | Alive check + timestamp |
| `/status` | Uptime, active calls, public URL |
| `/stats` | Call count, memory, TTS cache size |
| `/calls` | Last 5 call records with topics |
| `/debug` | Latency breakdown (avg from last 10 calls). Use `/debug -1`, `/debug -2` for per-call detail. |
| `/billings` | Check API balances (ElevenLabs chars, Twilio balance, OpenAI costs) |
| `/recording-on` | Start recording calls (Twilio recording) |
| `/recording-off` | Stop recording calls |
| `/restart` | Restart AVA (requires `/restart confirm`) |
| `/help` | Command list |

---

## Owner Instruction Injection

```mermaid
flowchart LR
    subgraph Signal["Owner sends via Signal"]
        A["tell him I'll call at 3"]
        B["ask for order number"]
        C["be more formal"]
        D["end"]
    end

subgraph GPT["AVA injects into GPT context"]
        A2["[RELAY_TO_CALLER: I'll call at 3]"]
        B2["[ASK_CALLER: order number]"]
        C2["[OWNER_INSTRUCTION: be more formal]"]
        D2["END_CALL_NOW + force_end flag"]
    end

A --> A2
    B --> B2
    C --> C2
    D --> D2

GPT --> Response["GPT acts on markers naturally within response"]

style Signal fill:#f0f9ff,stroke:#2563eb
    style GPT fill:#ecfdf5,stroke:#059669
```

---

## GPT Response Meta Block

Every GPT response ends with an invisible metadata block:

```
Hello, I'm Maya, Jacek's assistant. How can I help you today?

| Field | Purpose |
|-------|---------|
| `end_call` | `true` → AVA hangs up after this response |
| `urgency` | `low` / `medium` / `high` → emoji in Signal summary |
| `topic` | Short English description for Signal notifications |
| `caller_name` | First name if mentioned by caller |
| `lang` | Two-letter code (pl, en, de) → used to switch STT language |

---

## Docker Compose Services

```mermaid
graph LR
 subgraph compose["docker-compose.yml"]
 ava["ava FastAPI :8000 Python 3.11"]
 signal["signal-cli REST API :8080 Native mode"]
 caddy["caddy :80 / :443 Let's Encrypt"]
 tunnel["cloudflared Cloudflare Tunnel outbound only"]
 end

caddy -.-|"profile: caddy"| note1["Open ports 80/443"]
    tunnel -.-|"profile: tunnel"| note2["No open ports"]

style ava fill:#059669,color:#fff
    style signal fill:#2563eb,color:#fff
    style caddy fill:#d97706,color:#fff
    style tunnel fill:#7c3aed,color:#fff
```

---

## Environment Variables (complete reference)

| Variable | Default | Description |
|----------|---------|-------------|
| **Twilio** | | |
| `TWILIO_ACCOUNT_SID` | (required) | Twilio account identifier |
| `TWILIO_AUTH_TOKEN` | (required) | Auth token, also validates webhook signatures |
| `TWILIO_PHONE_NUMBER` | (required) | Your Twilio virtual number |
| **Signal** | | |
| `SIGNAL_CLI_URL` | `http://signal-cli:8080` | Internal signal-cli API address |
| `SIGNAL_SENDER_NUMBER` | (required) | Bot's Signal number |
| `SIGNAL_RECIPIENT` | (required) | Your personal Signal number |
| `SIGNAL_LANG` | `en` | Signal notification language (`en` / `pl`) |
| **LLM** | | |
| `OPENAI_API_KEY` | (required) | OpenAI API key |
| `LLM_PROVIDER` | `openai` | LLM backend: `openai` or `groq` |
| `LLM_MODEL` | auto | Model name (default: `gpt-4o-mini` for OpenAI, `llama-3.3-70b-versatile` for Groq) |
| `LLM_SUMMARY_MODEL` | auto | Model for call summaries (default: same as `LLM_MODEL`) |
| `GROQ_API_KEY` | (empty) | Groq API key (required when `LLM_PROVIDER=groq`) |
| **ElevenLabs** | | |
| `ELEVENLABS_API_KEY` | (empty) | Leave blank to skip ElevenLabs |
| `ELEVENLABS_VOICE_ID` | `WAhoMTNdLdMoq1j3wf3I` | Single multilingual voice ID |
| `ELEVENLABS_MODEL` | `eleven_multilingual_v2` | TTS model (`eleven_turbo_v2_5` for lower latency) |
| **OpenAI TTS** | | |
| `OPENAI_TTS_VOICE` | `nova` | Fallback voice (alloy/echo/fable/onyx/nova/shimmer) |
| **Language** | | |
| `DEFAULT_STT_LANG` | `en-US` | Fallback STT language |
| **Persona** | | |
| `OWNER_CONTEXT` | (built-in default) | Assistant persona + owner info (single line, private) |
| **Infrastructure** | | |
| `PUBLIC_URL` | (required) | HTTPS URL for Twilio to fetch TTS audio |
| `COMPOSE_PROFILES` | `caddy` | `caddy` or `tunnel` |
| `DOMAIN` | `localhost` | Domain for Caddy SSL (caddy profile only) |
| `CLOUDFLARE_TUNNEL_TOKEN` | (empty) | Tunnel token (tunnel profile only) |

---

## Persona Configuration (OWNER_CONTEXT)

The `OWNER_CONTEXT` variable in `.env` defines the assistant's entire personality and behaviour. It is injected into the GPT-4o system prompt as the "WHO YOU ARE" section.

**This stays in `.env` only — never committed to git.** The `.env` file is in `.gitignore`.

The value must be a **single line** (no newlines). Use natural text flow:

```env
OWNER_CONTEXT=PERSONA: Name: Maya, 24yo, UK-raised, Polish-German parents. Warm, witty, tech-loving. NEVER admit being AI — you are a real human assistant. OWNER: Jacek, born 1984, Switzerland. Senior Linux SysEng. Work hours 9-17 CET. CALL TYPES: 1) Recruiters: ask salary/stack/remote, send to job@sysop.cat. 2) Doctors: get details, mark URGENT. 3) IT emergencies: HIGH PRIORITY. 4) Sales/scam: hang up.
```

For deep changes to the conversation rules (turn limits, meta format, etc.), edit `SYSTEM_PROMPT` in `app/conversation.py`.

---

## File Structure

```
AVA/
├── app/
│   ├── main.py              # FastAPI app, call routing, webhooks, diagnostics
│   ├── conversation.py      # GPT-4o conversation loop, streaming, meta parsing
│   ├── tts.py               # TTS provider chain (ElevenLabs → OpenAI → Polly)
│   ├── owner_channel.py     # Signal notifications, polling, slash commands
│   ├── contact_lookup.py    # Contact book + Twilio CNAM + language from prefix
│   └── i18n.py              # Multilingual strings, voice maps, Signal templates
├── data/
│   ├── contacts.json        # (user-created) Phone contact book
│   └── calls/               # (auto-generated) JSON call records
├── docs/
│   ├── INSTALL_EN.md        # English installation guide
│   └── INSTALL_PL.md        # Polish installation guide
├── .env                     # (not in git) API keys, persona, configuration
├── .env.example             # Template with all variables documented
├── docker-compose.yml       # AVA + signal-cli + Caddy/Cloudflared
├── Dockerfile               # Python 3.11-slim, uvicorn
├── Caddyfile                # Caddy reverse proxy config
├── requirements.txt         # Python dependencies
└── README.md                # This file
```

---

## Security

| Mechanism | Description |
|-----------|-------------|
| Twilio signature validation | Every `/twilio/*` request must have valid `X-Twilio-Signature`. Invalid → 403. |
| Direct call rejection | Only forwarded calls are answered. Direct calls to the Twilio number are rejected (busy), unless the caller is in `contacts.json`. |
| Rate limiting | 30 requests/min per IP. Exceeding → 429. |
| Hidden app port | Port 8000 internal only. Traffic via Caddy HTTPS (:443) or Cloudflare Tunnel. |
| Signal sender filter | Only messages from `SIGNAL_RECIPIENT` are processed. Others are logged and ignored. |
| Audio file validation | Filenames must match `[a-f0-9]{32}\.mp3`. Path traversal blocked. |
| Security headers | Caddy adds HSTS, X-Frame-Options DENY, X-Content-Type-Options nosniff. |
| Disabled API docs | `/docs`, `/redoc`, `/openapi.json` endpoints are off. |

---

## Cost Estimate

| Service | Rate | Typical 2-min call |
|---------|------|--------------------|
| Twilio Voice | $0.013/min | ~$0.03 |
| Twilio STT (enhanced) | $0.02/15s | ~$0.16 |
| OpenAI Whisper | $0.006/min | ~$0.001 (first turn only) |
| OpenAI GPT-4o-mini | ~$0.0006/1k tokens | ~$0.001 |
| ElevenLabs | from $5/month | (30k chars free tier) |
| Twilio CNAM Lookup | $0.01/query | $0.01 (unknown numbers only) |

**Typical call: ~$0.20–0.25** (with GPT-4o-mini costs are significantly lower)

---

## Signal Commands

### During a call

| Message | What happens |
|---------|--------------|
| `tell him I'll call back tomorrow at 10` | AVA naturally relays this to the caller |
| `ask for the order number` | AVA asks the caller |
| `end` / `stop` / `koniec` | AVA wraps up the call gracefully |
| `status` or `?` | Confirms whether a call is active |
| Any other text | Forwarded as a generic instruction |

---

## Setup

See the detailed installation guides:
- **English**: [docs/INSTALL_EN.md](docs/INSTALL_EN.md)
- **Polish**: [docs/INSTALL_PL.md](docs/INSTALL_PL.md)

### Quick start

```bash
cp .env.example .env
# Edit .env — fill in API keys, OWNER_CONTEXT, PUBLIC_URL
mkdir -p data/calls
docker compose up -d
curl https://your-domain.com/health
```

---

## Troubleshooting

```bash
# Twilio can't reach the webhook?
curl -I https://your-domain.com/health

# TTS audio not playing?
docker compose logs ava | grep -i tts

# Signal not sending?
docker compose logs ava-signal-cli
curl http://localhost:8080/v1/accounts

# Check active calls
# Send "status" or "/status" to the Signal bot

# Clear TTS cache (after voice change)
docker exec ava sh -c 'rm -f /tmp/tts_cache/*.mp3'

# View recent call logs
ls -lt data/calls/ | head
```

ava

ava Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

ava

ava Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README