内容
<div align="center">
<h1>SARA</h1>
<p><strong>Open-source WhatsApp AI agent — 20 verticals defined, 13 brains included</strong></p>
[](https://opensource.org/licenses/Apache-2.0)
[](https://hub.docker.com)
[](https://www.typescriptlang.org)
[](CONTRIBUTING.md)
[](https://github.com/Alessandro114/sara)
**SARA is a self-hostable AI agent for WhatsApp that understands your industry. 13 industry brains included in this repo; 20 verticals are defined in the open [`scala-agent-definitions`](https://github.com/Alessandro114/scala-agent-definitions) schema (Apache-2.0), and you can write your own. 83 tool definitions. Function calling. RAG. Multi-language.**
[Documentation](#configuration) · [Cloud Version](https://get-scala.com) · [Contributing](CONTRIBUTING.md)
</div>
---
## Why SARA?
- **vs Generic chatbots (Tidio, Landbot):** SARA knows your industry. A restaurant bot books tables and checks allergens. A clinic bot schedules patient appointments. Generic bots just answer FAQs.
- **vs Enterprise platforms (Salesforce, ServiceNow):** Self-host for free. No vendor lock-in. No six-figure contracts. Full source access.
- **vs Building from scratch:** the verticals are already wired — knowledge bases, tools, prompts, fallbacks. 13 brains ship in this repo, 20 are defined in the open schema. You add your data, not your architecture.
- **vs Other open source (Chatwoot, Typebot):** SARA is an AI agent with function calling, RAG retrieval, PII anonymization, and autonomy levels — not just a chat widget with canned replies.
---
## Features
### AI Engine
- Multi-provider LLM chain (Groq → Cerebras → SambaNova → Mistral) with automatic failover — zero downtime when one provider is rate-limited
- RAG retrieval: PostgreSQL + pgvector (1024d embeddings) + BM25 keyword fallback for hybrid search
- Function calling with 83 tool definitions across 20 industry verticals — agents can actually *do* things, not just respond
- PII anonymization before every external LLM call — names, phones, emails are masked at the boundary
- Prompt injection guardrails: 40+ patterns, 4 languages, evaluated before every turn
- Hallucination detection + confidence scoring on AI responses
### WhatsApp
- Voice messages: speech-to-text via Whisper
- Image understanding via vision AI
- PDF and document processing
- Location, contacts, and sticker support
- Rate limiting + anti-ban protection built in
### Industry Verticals (20)
| Vertical | Key tools |
|---|---|
| DineOS (restaurants) | Table booking, menu, allergens, price list |
| PropertyOS (real estate) | Listings, valuations, inspections |
| BeautyOS (salons) | Appointments, treatments, membership |
| MotorOS (automotive) | Test drives, service booking, trade-in |
| WellnessOS (gyms) | Class booking, membership management |
| PraxisOS (clinics) | Patient scheduling, intake forms |
| ShopOS (retail) | Product search, stock check, reservations |
| AgencyOS | Client intake, brief collection |
| StudioOS | Creative project booking, SAL generation |
| CleanOS | Cleaning service scheduling |
| TravelOS | Itinerary search, booking requests |
| NetworkOS | Event registration, networking |
| DermalyOS | Dermatology intake, appointment |
| FranchiseOS | Franchise lead qualification |
| ProjectOS | Project status, task updates |
| ReputationOS | Review collection, feedback routing |
| LandIQ | Real estate feasibility queries |
| AdOS | Campaign briefing, performance queries |
| ServiceOS | Field service scheduling |
| TenderOS (public tenders) | Tender parsing, deadlines, gap analysis, checklists |
### Platform
- Multi-language: Italian, English, Spanish, Portuguese
- Conversation memory: history (L4) + contact profiles (L5) + RAG knowledge base (L6)
- Human takeover: AI / human / hybrid mode per conversation, switchable in real time
- Built-in dashboard: analytics, session overview, conversation logs
- Autonomy levels: OFF / OBSERVE / SEMI-AUTO / FULL-AUTO configurable per tenant
---
## Quickstart
```bash
git clone https://github.com/Alessandro114/sara.git
cd sara
cp .env.example .env
# Edit .env — minimum required: GROQ_API_KEY (free at console.groq.com)
docker compose up -d
```
- Verify: `curl http://localhost:3006/api/sara/health`
- Connect WhatsApp: open `http://localhost:3006/api/sara/qr` and scan the QR code with your WhatsApp number.
SARA is running. Send a message to the connected number.
---
## Architecture
```
Customer (WhatsApp)
|
v
WAHA (WhatsApp Web API)
|
v
SARA Core (Node.js + Fastify)
|-- AI Provider Chain (Groq -> Cerebras -> SambaNova -> Mistral)
|-- RAG Engine (PostgreSQL + pgvector, 1024d)
|-- Vertical Brain (DineOS / PropertyOS / BeautyOS / ...)
|-- Function Calling (book_table / search_listings / schedule_appointment / ...)
|-- PII Anonymizer
|-- Prompt Injection Guard
|-- Memory (L4 history + L5 profiles + L6 RAG)
|
v
Response -> Customer
```
---
## Configuration
Key environment variables (see `.env.example` for the full list):
| Variable | Required | Description |
|---|---|---|
| `GROQ_API_KEY` | Yes | Primary LLM provider. Free at [console.groq.com](https://console.groq.com) |
| `CEREBRAS_API_KEY` | No | Fallback LLM provider |
| `SAMBANOVA_API_KEY` | No | Fallback LLM provider |
| `MISTRAL_API_KEY` | No | Fallback LLM provider |
| `DATABASE_URL` | Yes | PostgreSQL connection string (pgvector required) |
| `WAHA_URL` | Yes | WAHA instance URL (default: `http://localhost:3004`) |
| `WAHA_API_KEY` | Yes | WAHA authentication key |
| `VERTICAL` | Yes | Active vertical slug (e.g. `dineos`, `propertyos`, `beautyos`) |
| `EMBED_MODEL` | No | Embedding model — must be 1024d. Default: `mxbai-embed-large` |
| `OLLAMA_BASE_URL` | No | Ollama URL if using local embeddings |
| `SARA_AUTONOMY_LEVEL` | No | `OFF` / `OBSERVE` / `SEMI_AUTO` / `FULL_AUTO`. Default: `OBSERVE` |
> The embedding model **must produce 1024-dimensional vectors**. Supported:
> `bge-m3`, `mxbai-embed-large`, `jina-embeddings-v3`. Do not use
> `nomic-embed-text` (768d) — a dimension mismatch is rejected by pgvector at
> the first insert, so you find out immediately.
>
> **The dangerous case is the opposite one: two different models that produce
> the SAME dimension.** All three supported models emit 1024d, so pgvector
> accepts vectors from any of them into the same column without a single
> error — but they are different vector spaces, and the distance between a
> vector from one model and a vector from another means nothing. A mixed index
> gives no symptom: it returns plausible, effectively random results. With an
> HNSW index it is worse than imprecise, because the vectors of the
> minority model become unreachable rather than merely mis-ranked.
>
> The failure mode to watch for: a second process — a different importer, a
> background job, a migration script — writing into the same table with its own
> model setting. Nothing errors, and the corpus degrades silently.
>
> **If you change model, re-embed the whole corpus** — never incrementally —
> and store the model name on every row so a mixed index is detectable at all:
>
> ```sql
> SELECT embedding_model, count(*) FROM your_table GROUP BY 1;
> ```
>
> Without that column, a mixed index is indistinguishable from a healthy one.
---
## Self-hosted vs Cloud
| Feature | Self-hosted (free) | Cloud ([get-scala.com](https://get-scala.com)) |
|---|---|---|
| AI chat engine | Yes | Yes |
| 20 verticals defined / 13 brains included | Yes | Yes |
| Function calling (30+ tools) | Yes | Yes |
| RAG engine | Yes | Yes |
| Multi-language | Yes | Yes |
| Dashboard | Yes | Yes |
| Multi-number (multi-tenant) | No | Yes |
| White-label | No | Yes |
| Advanced analytics | No | Yes |
| Priority support | No | Yes |
| Managed infrastructure | No | Yes |
---
## Contributing
Contributions are welcome. See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
Good first issues are labeled [`good first issue`](https://github.com/Alessandro114/sara/labels/good%20first%20issue). If you want to add a new vertical, open an issue first so we can align on the tool schema before you build.
---
## License
SARA is licensed under [Apache 2.0](LICENSE).
You can self-host it, modify it, and ship it inside a commercial or closed-source product. No copyleft obligation: you are not required to publish your changes. Keep the `LICENSE` and `NOTICE` files with any copy you distribute, and state what you changed. Apache 2.0 also grants you a patent licence from the contributors.
If you want commercial support, a hosted deployment, or the industry packs that are not in this repo, write to [ale@get-scala.com](mailto:ale@get-scala.com).
---
## Built by
SARA is the AI engine behind [S.C.A.L.A. AI OS](https://get-scala.com) — an enterprise AI operating system for businesses.
Built by [Alessandro Binda](https://www.linkedin.com/in/alessandrobinda/), a GM turned AI builder with 18 years of industry experience and P&L ownership up to EUR 33M.
- Website: [get-scala.com](https://get-scala.com)
- LinkedIn: [linkedin.com/in/alessandrobinda](https://www.linkedin.com/in/alessandrobinda/)
- Twitter/X: [@alessandrobinda](https://x.com/alessandrobinda)
---
<div align="center">
**If SARA saves you time, [give it a star](https://github.com/Alessandro114/sara)**. It helps other developers discover it.
</div>
---
## Ecosystem
Part of the **S.C.A.L.A.** open-source ecosystem:
| Project | What it does |
|---------|-------------|
| [SARA](https://github.com/Alessandro114/sara) | WhatsApp AI agent — 20 verticals defined, 13 brains included |
| [LandIQ](https://github.com/Alessandro114/landiq) | Autonomous real estate feasibility agent |
| [scala-sites](https://github.com/Alessandro114/scala-sites) | 100 vertical website templates (Next.js, MIT) |
| [scala-agent-definitions](https://github.com/Alessandro114/scala-agent-definitions) | 80 AI tool definitions for 20 verticals |
| [scala-mcp-server](https://github.com/Alessandro114/scala-mcp-server) | MCP server for Claude/ChatGPT — 250M+ companies |
| [Score SDKs](https://github.com/Alessandro114/scala-score-js) | Company data — [JS](https://npmjs.com/package/scala-score) · [Python](https://pypi.org/project/scala-score) · [Go](https://github.com/Alessandro114/company-lookup-go) · [Rust](https://github.com/Alessandro114/score-rust) · [Deno](https://github.com/Alessandro114/scala-score-deno) |
| [enrich-companies](https://github.com/Alessandro114/enrich-companies) | CSV enrichment CLI — [npm](https://npmjs.com/package/enrich-companies) · [pip](https://pypi.org/project/enrich-companies) |
| [n8n node](https://github.com/Alessandro114/n8n-nodes-scala) | n8n community node for company data |
| [awesome-whatsapp-ai](https://github.com/Alessandro114/awesome-whatsapp-ai) | Curated list of WhatsApp AI tools & frameworks |
连接信息
猜你喜欢
暂无推荐服务器