Table of contents
We started GreenPT with a simple conviction: powerful AI doesn’t have to come at the planet’s expense. Today we’re thrilled to turn that conviction into a product you can use. Meet Frida — an AI-powered customer support platform that helps organizations deliver exceptional visitor and customer experiences while keeping its environmental footprint as small as possible.
Frida is not another chatbot builder. It’s a full-stack platform that combines intelligent agents, a state-of-the-art retrieval pipeline, real integrations with the tools you already use, and transparent environmental impact tracking — all wrapped in a product that’s genuinely easy to set up.
In this post, we’ll walk you through everything Frida can do, why we built it the way we did, and how you can get started for free today.
The Problem We Set Out to Solve
Every day, support teams field the same questions dozens — sometimes hundreds — of times. “What are your opening hours?” “How much does a ticket cost?” “Do you offer group discounts?” “Where can I park?” Meanwhile, visitors leave websites without the answers they need simply because no one was there to help. For museums, galleries, cultural institutions, e-commerce stores, and countless other organizations, this is a real and growing challenge.
The promise of AI chatbots isn’t new. But most of the tools on the market fall into one of two camps: overly simplistic bots that can only parrot pre-written FAQ answers, or enterprise platforms that take months to configure and cost a fortune. Neither option is great if you’re a mid-size organization that just wants helpful, accurate answers for your visitors — without a dedicated AI engineering team.
And then there’s the sustainability question. Every AI response consumes energy and generates carbon emissions. The industry’s current approach is to pretend this cost doesn’t exist. We believe organizations deserve to know exactly what their AI usage costs the planet, and to have the option to minimize that impact without sacrificing quality.
Frida bridges all of these gaps.
Meet the AI Agents
At the heart of Frida are Agents — AI-powered assistants that you can tailor to your organization in minutes. Each agent has its own name, personality, instructions, and knowledge base. This isn’t a one-size-fits-all chatbot; you decide exactly how your agent should behave.
Personality & Instructions
When you create an agent, you write a set of instructions that tell it who it is and how it should respond. For example, you might tell a museum agent: “You are a friendly and knowledgeable guide for the Rijksmuseum. Always answer in a warm, welcoming tone. If you don’t know the answer, suggest the visitor contact our front desk.”
These instructions shape every response the agent produces. You can be as specific or as general as you like. Want the agent to always respond in the visitor’s language? Add it to the instructions. Want it to avoid discussing certain topics? Say so.
Choose Your Model
Frida gives you control over the AI model that powers each agent. Our recommended default is GreenR — GreenPT’s own reasoning model, specifically optimized for the best balance of answer quality and environmental efficiency. But if your use case demands peak performance from frontier models, you can switch to GreenL (our large model), open-source options like Llama 3.3 70B, Mistral Nemo, DeepSeek R1, or Gemma 3 27B, or even frontier models like Claude Sonnet 4.5, Claude Opus 4.5, and Mistral Large 3 via EU Router.
You choose the model per agent, so you can have one agent running on GreenR for everyday questions and another on Claude for complex reasoning tasks — all within the same organization.
Multi-Agent Architecture
Here’s where things get interesting. Frida supports assigning multiple agents to a single channel, with a drag-and-drop priority system. The first agent in the list is the Primary Agent — its name, personality, and instructions drive the responses. But when the system searches for relevant context, it searches the knowledge bases of all assigned agents.
This opens up powerful patterns. You can create separate specialist agents — a Product Expert, a Billing Assistant, a Technical Support agent — each with its own knowledge base and tailored instructions. Assign all three to one channel, and the primary agent will draw on the combined knowledge of all specialists when composing answers.
Why does this matter? Because organized, domain-specific knowledge helps the AI find relevant context faster and produce more accurate responses. Instead of one giant agent with everything thrown into it, you get a well-organized team of specialists working together.
—
Knowledge Paths: Teach Your Agent Anything
Knowledge Paths are the building blocks of what your agent knows. Think of them as the training material — the facts, documents, and information that your agent draws on when answering questions.
Multiple Content Types
You can create Knowledge Paths from a variety of sources:
- Text content — Write or paste information directly.
- URL import — Point Frida at any web page and it will extract the content automatically.
- Document upload — Upload PDFs, Word documents, spreadsheets, and presentations. Frida uses Docling, a document intelligence engine, to convert complex file formats into clean, structured text.
- Integration import — Pull content directly from WordPress, WooCommerce, Shopify, Google Drive, OneDrive, Notion, or Linear (more on integrations below).
Each Knowledge Path can include images and videos alongside text, enabling richer, more visual responses when your agent references that material.
Organize with Tags
Knowledge Paths support tags, which help you organize content into categories. Tags are also used by the Knowledge Browser feature (covered below) to group articles for visitors who prefer to browse rather than chat.
Deep Research
For Knowledge Paths that need comprehensive, well-sourced content, Frida includes a Deep Research feature. Enter a research query, and Frida will search the web using GPT Researcher, gather sources, analyze them, and then use GreenR to synthesize a comprehensive research report — complete with citations. The entire process streams in real-time, so you can watch sources being discovered and the report being written live.
Once the research is complete, you review and edit the report before saving it. Nothing gets published to your agent’s knowledge base without your approval.
Deploy Everywhere Your Audience Is
One of Frida’s core design principles is that your visitors should be able to reach your AI agent on whatever channel they already use. Setting up your agent once and deploying it everywhere shouldn’t require rebuilding anything.
Website Widget
The website widget is the most common deployment channel. It’s a customizable chat bubble that you embed on your site with a single line of JavaScript. But “customizable” barely begins to describe it.
Widget Templates. Choose from pre-built themes — Classic (a customizable sage green theme) or Lucio (available in light and dark variants) — and fine-tune from there.
Full Appearance Control. Every color in the widget is configurable: header background, text colors, user and reply bubble colors, input field styling, footer appearance, source link colors, and more. Choose from a curated set of Google Fonts to match your brand typography.
Welcome Screen. Before the conversation starts, visitors see a polished welcome screen with your agent’s name, an introduction, and a “Start Conversation” button. You can customize every element — the title, subtitle, agent intro text, button label, and even add contact options (email, phone, links) that appear alongside the chat.
Quick-Start Buttons. Help visitors get started with pre-configured suggestion buttons like “What are your opening hours?” or “Tell me about group discounts.” These appear at the top of the chat and give visitors an easy entry point.
Proactive Message Balloon. Don’t wait for visitors to click the chat button. Configure a proactive message balloon that appears after a customizable delay (e.g., 5 seconds) near the widget button. Add a notification sound (pop, chime, or bell) to draw attention. It’s a subtle, elegant way to say “Hey, I’m here if you need help” — and it significantly increases engagement.
Knowledge Browser. Enable a “Browse” tab alongside the chat, letting visitors explore your knowledge base directly. Articles are organized by category with full-text search. Visitors can switch between asking questions and reading articles, all within the same widget.
Visibility Hours. Control when the widget is online. Set specific hours per day of the week — for example, Monday through Friday, 9 AM to 5 PM in your timezone. Outside those hours, visitors see an offline message instead of the chat. Perfect for organizations that want to align AI availability with staff availability for escalations.
Voice Input. Visitors can speak their questions instead of typing. Frida uses GreenPT’s speech-to-text model (Green-S Pro) for real-time transcription via WebSocket streaming, with automatic fallback to server-side transcription or the browser’s native Web Speech API. Users select their language before recording, with support for 13+ languages including English, Dutch, German, French, Spanish, Italian, Portuguese, Polish, Russian, Chinese, Japanese, Korean, and Arabic.
Accessibility. The widget is built with accessibility in mind — ARIA labels, keyboard navigation, focus traps in modals, and optional sound cues for new messages and typing indicators.
Connect your Frida agent to WhatsApp Business and respond to customer messages automatically. WhatsApp is the world’s most popular messaging app, with over 2 billion users. For organizations that serve international audiences, this channel alone can transform their support experience.
Microsoft Teams
Deploy your agent inside your organization’s Microsoft Teams environment. Team members can chat with the bot in personal conversations or @mention it in team channels. This is particularly powerful for internal support scenarios — IT help desks, HR FAQ bots, or knowledge base assistants that your staff can query directly in the tools they already use.
Setting up Teams requires an Azure Bot registration, but Frida provides a comprehensive step-by-step guide that walks you through every screen, from creating the Azure resource to uploading the app package.
Slack
Bring your agent into Slack and let it respond to direct messages. If your team already lives in Slack, this is the fastest way to make organizational knowledge accessible to everyone — without leaving the workspace.
Discord
For community-driven organizations — gaming communities, open-source projects, educational programs — Discord is often the primary communication hub. Deploy your Frida agent as a Discord bot and let it serve your community 24/7 with always-on, knowledge-backed answers.
Integrations That Actually Work
Frida doesn’t exist in a vacuum. Your organization’s knowledge lives in WordPress, in Google Drive, in Notion, in Shopify. Frida connects to all of these — and more.
WordPress & WooCommerce
This is one of our deepest integrations. Frida offers two paths:
The WordPress Plugin. Install the “Frida by GreenPT” plugin from the WordPress Plugin Directory. Activate it, click “Connect to Frida” in the sidebar, and a guided popup walks you through connecting your account, importing your posts and pages, creating an agent, and deploying a widget — all in one flow. The plugin automatically embeds the chat widget on every page of your WordPress site.
API-Based Integration. For more advanced setups, connect via the Integrations page. Frida auto-detects WordPress REST API endpoints, discovers your posts and pages, handles multilingual content (WPML and Polylang are both supported), extracts Yoast SEO metadata and Advanced Custom Fields for richer context, and even detects product data if you’re running WooCommerce.
Business plan users get automatic syncing — Frida periodically re-imports your content so your agent always has the latest information.
Shopify
Connect your Shopify store to import your product catalog. Frida extracts product titles, descriptions, variants, pricing, images, and collection information. You can also enable live product lookup — a Chat Tool that lets the agent search your Shopify inventory in real-time during conversations, so pricing and stock information is always current.
Google Drive & OneDrive
Connect your cloud storage and select specific folders or files to import. Frida extracts text from documents, spreadsheets, and presentations and converts them into Knowledge Paths. This is especially powerful for organizations that maintain documentation in Google Docs or Word — you’re effectively turning your existing documentation into an AI knowledge base with zero rewriting.
Notion
If your team uses Notion for wikis, documentation, or knowledge management, Frida can sync your pages and databases directly. Select which Notion pages to share during the connection flow, and they’ll become part of your agent’s knowledge.
Linear
For engineering and support teams using Linear for issue tracking, this integration lets your agent reference open issues, project documentation, and known bugs. Your support agent can look up ticket statuses and help users find relevant issues without leaving the chat.
Sitemap Import
Don’t have a CMS integration? No problem. Point Frida at any website’s sitemap.xml and it will discover all pages, let you select which ones to import, and create Knowledge Paths from the crawled content. This works with any website that has a sitemap — which is most of them.
Under the Hood: The RAG Pipeline
We’re proud of the engineering behind Frida’s answer generation. When a visitor asks a question, they don’t get a generic GPT response — they get a carefully researched answer drawn from your specific knowledge base. This is possible because of our multi-stage Retrieval Augmented Generation (RAG) pipeline.
How It Works
“`
User Query → Query Rewriting → Embedding → Hybrid Search → Reranking → Context Selection → LLM Response
“`
1. Semantic Chunking. When documents are ingested, content is intelligently chunked to preserve semantic meaning. The system respects Markdown headers, code blocks, tables, and lists. Each chunk carries hierarchical context — for example, [Getting Started > Installation] — so the AI knows where in the document the information comes from.
2. Query Rewriting. When a visitor asks a question, the system can optionally rewrite the query to improve retrieval. This is particularly useful for follow-up questions that reference earlier context (e.g., “What about the price?” after asking about a specific product).
3. Hybrid Search. Frida doesn’t rely on a single search strategy. It combines semantic vector search (using pgvector embeddings) with full-text search (using PostgreSQL’s tsvector). The results are blended using configurable weights (default: 70% vector, 30% full-text). This hybrid approach catches both semantically similar content and exact keyword matches.
4. Reranking. The initial search results are reranked using GreenPT’s reranking model to push the most relevant chunks to the top. This dramatically improves answer quality by ensuring the LLM sees the best context first.
5. Context Selection. A configurable context budget (default: 12,000 characters) determines how much retrieved content is passed to the LLM. This prevents context window overflow while ensuring the model has enough information to generate a thorough answer.
6. Response Generation. Finally, the selected context — along with your agent’s instructions and the conversation history — is passed to the chosen AI model, which generates a natural, contextual response.
The result? Fewer hallucinations, more accurate answers, and visitors who actually get the information they need. The system can even cite its sources, showing visitors links to the original pages or documents where the information came from.
Configurable at Every Stage
Every stage of the pipeline is tunable via environment variables. Want to increase the number of reranked chunks? Adjust the context budget? Change the vector-to-fulltext weight ratio? Tweak the relevance score threshold? You can. We’ve set sensible defaults, but the knobs are there for teams that want to optimize.
Built for Sustainability
This is where Frida truly stands apart. Every AI response consumes energy and generates carbon emissions. Most platforms ignore this entirely. We don’t.
Eco-Optimized Models
Our default models — GreenR (reasoning) and GreenL (large) — are specifically optimized for the best balance of quality and environmental efficiency. GreenR is our recommended model and comes selected by default for every new agent, so you’re making the greener choice without even thinking about it.
But “eco-optimized” doesn’t mean “less capable.” GreenR supports full chain-of-thought reasoning with a dedicated thinking mode. It can work through multi-step problems, debug logic, and handle complex analytical questions — just like frontier reasoning models, but with a fraction of the environmental footprint.
Transparent Impact Tracking
This is something we’re genuinely proud of. Frida tracks the energy consumption (in watt-milliseconds) and CO₂ emissions (in micrograms) of every single AI interaction that uses our proprietary models. This data is surfaced in two places:
The Widget. Visitors can see the environmental impact of their conversation directly in the chat footer. It’s a small but meaningful way to demonstrate your organization’s commitment to sustainability.
The Insights Dashboard. Organization admins get a dedicated impact metrics panel showing total CO₂ emissions, total energy consumption, averages per message, and the number of interactions with impact data. The numbers are formatted in human-readable units — micrograms become milligrams become grams as your usage grows.
No guesswork. No greenwashing. Just real data about the environmental cost of your AI usage.
Frontier Models When You Need Them
We’re realistic about the fact that some use cases demand the absolute best performance from the latest models. For those scenarios, Frida offers access to frontier models through EU Router — a European AI gateway that provides access to models from Anthropic, Mistral, Moonshot, and Qwen, all routed through EU-based infrastructure.
Available frontier models include:
- Mistral Large 3 — Mistral’s most capable model
- Claude Sonnet 4.5 — Anthropic’s balanced model
- Claude Opus 4.5 — Anthropic’s most powerful model
- Qwen3 Coder 480B — Large code-specialized model
- Kimi K2 — Moonshot’s instruction-following model
Bring your own API key, and switch between eco-optimized and frontier models on a per-agent basis. This way, you can use GreenR for 95% of interactions and reserve frontier models for the 5% that truly need them.
Privacy-First, GDPR-Compliant
Frida is designed with privacy as a fundamental requirement, not an afterthought.
Encryption. All data is encrypted in transit (TLS) and at rest. Sensitive credentials like API keys are encrypted before being stored in the database.
Data isolation. Multi-tenant architecture ensures every organization’s data is completely isolated. Database queries are scoped by organization ID at every level, with security checks built into every storage method.
No training on your data. Your knowledge base content is never used to train external AI models. Your data is your data.
EU infrastructure. EU Router ensures that frontier model requests are routed through EU-based infrastructure, keeping your data within European borders for full GDPR compliance.
Origin validation. The widget validates request origins against your configured allowed origins list, preventing unauthorized embedding of your chat widget on third-party sites.
Input sanitization. All user inputs are sanitized at multiple levels — HTML escaping, prompt injection guards, length limits, and content type validation — to prevent XSS, injection attacks, and abuse.
Quality Monitoring & Feedback
Deploying an AI agent is just the beginning. You need to know how it’s performing and where it’s falling short. Frida gives you the tools to monitor quality and continuously improve.
Flagged Messages
Visitors can flag any AI response they find incorrect, unhelpful, inappropriate, or otherwise problematic. Each flag includes a reason and optional details. Flagged messages appear in a dedicated admin dashboard where you can review the full conversation context, add notes, and update the status (New → Reviewing → Resolved or Dismissed).
This is the fastest feedback loop for improving your agent: see exactly which responses aren’t working, understand why, and update your agent’s instructions or knowledge base accordingly.
Conversation Ratings
At the end of a conversation, visitors can rate their experience on a 1-to-5 star scale with an optional comment. These ratings flow into a Feedback Dashboard with statistics, distribution charts, channel breakdowns, and recent feedback entries. It’s the high-level pulse check that tells you whether your agent is meeting expectations.
Insights Dashboard
Beyond feedback, the Insights page gives you quantitative metrics: total conversations, total messages, average messages per conversation, active agents, channel breakdowns, agent breakdowns, and — of course — environmental impact metrics. Filter by date range to spot trends and track improvements over time.
Multilingual by Design
Frida is built for international audiences from the ground up.
Dashboard languages. The admin dashboard is fully translated into 9 languages: English, Dutch, French, Spanish, Italian, German, Swedish, Norwegian, and Portuguese. The system auto-detects your browser language and applies the right translation.
Widget languages. The embeddable widget supports the same 9 languages, with every label, placeholder, and message fully localized. You can set the widget language per channel, so a Dutch museum can have a Dutch widget while its English sister site has an English one.
Voice input languages. Speech-to-text supports 13+ languages including English, Dutch, German, French, Spanish, Italian, Portuguese, Polish, Russian, Chinese, Japanese, Korean, and Arabic.
AI response languages. The AI models themselves can understand and respond in 50+ languages. Your agent automatically detects the visitor’s language and responds accordingly. A Japanese visitor gets Japanese answers; a Brazilian visitor gets Portuguese answers — all from the same knowledge base.
The Playground: Test Before You Deploy
Before going live, you want to be confident your agent handles questions well. That’s what the Playground is for.
Select an agent and channel, then have a conversation with your agent just as a visitor would. Test different scenarios — common questions, edge cases, unusual phrasings, other languages. Adjust your agent’s instructions based on the results. Iterate until you’re satisfied.
The Playground is also invaluable for onboarding new team members. Let them interact with the agent to understand how it works and what it can do, without ever touching production.
Pricing That Grows With You
We believe every organization should be able to try AI-powered support without a credit card or a sales call.
Free
Build and test your AI agent, deploy a website widget, connect to WhatsApp, Microsoft Teams, Slack, and Discord. Full analytics, environmental insights, and user feedback. This is a real free tier — not a 14-day trial.
Individual
Everything in Free, plus WordPress and Sitemap content import. More agents, more channels, more Knowledge Paths.
Pro
Everything in Individual with expanded limits and access to EU Router for frontier models.
Business
Everything in Pro, plus automatic WordPress and Sitemap syncing, custom footer text, team management, website crawling, and the highest resource limits.
All plans support monthly or yearly billing (with a discount for yearly). Upgrade or downgrade anytime. No lock-in, no surprises.
Get Started Today
Setting up Frida takes less than five minutes.
- Create an account at frida.greenpt.ai — no credit card required.
- Build your first agent — give it a name, write instructions, choose a model, and set its personality.
- Add knowledge — upload documents, paste URLs, connect an integration, or run a Deep Research session.
- Deploy a channel — grab the embed code for your website, or connect to WhatsApp, Teams, Slack, or Discord.
- Test in the Playground — make sure everything works before going live.
- Monitor & improve — use flagged messages, ratings, insights, and environmental metrics to continuously refine your agent.
That’s it. Your visitors get instant, intelligent, around-the-clock support — and you get peace of mind knowing it’s powered by sustainable AI.
What’s Next
This launch is just the beginning. We have a lot planned for Frida in the months ahead:
- Real-time conversations viewer — A redesigned conversations page with WebSocket-powered live updates, advanced filtering, and conversation analytics.
- More integrations — We’re always adding new data connectors and Chat Tools based on what our users need.
- Enhanced analytics — Deeper insights into agent performance, topic clustering, and conversation patterns.
- Expanded model ecosystem — As new models become available, we’ll add them to the platform.
We’re building Frida in the open, informed by the feedback of the organizations that use it every day. If there’s something you need, tell us — we’re listening.
