Yet Another Agentic Web Chat (YAAWC)

Because the world definitely needed one more AI-powered search engine. You're welcome.

YAAWC (Pronounced: "yawck" — as in the sound you make when yet another AI search engine appears.) is an open-source, self-hosted, agentic AI search engine that actually reads web pages, manages research plans, spawns sub-agents for deep dives, and cites its sources — all while letting you pick from a buffet of LLM providers. It uses SearXNG under the hood so your queries stay private and the results stay fresh.

Table of Contents

• Why Does This Exist?
• Features at a Glance
• Focus Modes
• Agent Tools
• Deep Research (Sub-Agents)
• Dashboard Widgets
• LLM Providers
• Personalization & Personas
• Installation
  • Docker (Recommended)
  • Manual Setup
  • Ollama Connection Errors
• Updating
• Using as a Browser Search Engine
• API
• Network & Reverse Proxy
• Observability
• Contributing
• Acknowledgements

Why Does This Exist?

Most AI search tools either phone home with your data, cost a subscription, or give you yesterday's answers. YAAWC is fully open source, runs on your hardware, talks to whichever LLM you point it at, and searches the live web through a self-hosted SearXNG instance. It doesn't just retrieve links — it reads pages, extracts the good parts, ranks them by semantic similarity, and writes you a cited answer.

Want to know more about the architecture? See docs/architecture/README.md.

Features at a Glance

Category	Highlights
Agentic Search	LangGraph React agent with tool use, research planning, and multi-step reasoning
Deep Research	Spawns focused sub-agents that search → read → refine → search again
9 Agent Tools	Web search, URL summarization, image search & analysis, YouTube transcripts, PDF loading, file search, deep research, todo lists
10 LLM Providers	OpenAI, Anthropic, Groq, Ollama, Gemini, DeepSeek, LM Studio, OpenRouter, AI/ML API, Custom OpenAI
6 Embedding Providers	OpenAI, Ollama, Gemini, Xenova Transformers (local), AI/ML API, LM Studio
Dashboard Widgets	AI-powered info widgets with auto-refresh, drag-and-drop layout, export/import
Personas	Custom system prompts with built-in templates (scholarly, conversational, etc.)
Personalization	Per-message location and profile context injection
Privacy	Self-hosted SearXNG — no tracking, no data brokering, no "we updated our privacy policy" emails
Browser Integration	OpenSearch XML, autocomplete, ?q= URL queries with saved preferences
Streaming UI	Real-time tool calls, sub-agent progress, todo widgets, thinking/reasoning display
Image & Video Search	Dedicated search with gallery views and video embeds
File Research	Upload documents and research them with cited excerpts
Respond Now	Interrupt ongoing retrieval and get an immediate answer from what's been gathered so far
Model Visibility	Admins can hide models from the UI to prevent accidental usage
Dual Model Architecture	Separate Chat and System models, linkable or independent

Focus Modes

Switch modes at any time during a conversation:

Mode	Description	Tools
Web Search	Full agentic search across the internet	All 9 tools
Chat	Creative conversation — no web searching, no tools	None
Local Research	Research uploaded files with semantic search and citations	File search

Firefox AI prompts are auto-detected and handled conversationally.

Agent Tools

The LangGraph agent has access to the following tools (individually toggleable per conversation):

Tool	What It Does
Web Search	Queries SearXNG, retrieves top results, re-ranks by embedding similarity. Supports site: filters.
URL Summarization	Fetches a URL's content (via Readability/Cheerio/Playwright) and summarizes it or uses it directly.
Image Search	Searches for images via SearXNG (Bing Images, Google Images).
Image Analysis	Fetches an image and analyzes it using a vision-capable LLM (PNG, JPEG, GIF, WebP up to 10 MB).
YouTube Transcript	Retrieves the full transcript from a YouTube video.
PDF Loader	Extracts and returns content from a PDF URL.
File Search	Semantic similarity search across uploaded documents with configurable threshold.
Deep Research	Spawns a focused sub-agent for comprehensive multi-source investigation (see below).
Todo List	Manages a visible research plan (up to 10 tasks) with live progress in the UI.

Deep Research (Sub-Agents)

When the agent encounters a question that needs serious digging, it can invoke the Deep Research tool, which spawns an independent sub-agent with its own system prompt and tool access (web_search, url_summarization, image_search, youtube_transcript, pdf_loader — but not deep_research, because infinite recursion is nobody's friend).

Sub-agent progress streams live to the UI: task description, nested tool calls, and the final synthesized response — all collapsible and inspectable.

Dashboard Widgets

The /dashboard page provides a configurable grid of AI-powered widgets:

• Fetch content from web pages or HTTP endpoints
• Process fetched content with an AI prompt using any configured provider/model
• Drag, drop, and resize widgets on a responsive grid
• Auto-refresh at configurable intervals (minutes/hours)
• Export/import dashboard configurations as JSON
• Date/time template variables for dynamic prompts ({{current_utc_datetime}}, {{current_local_datetime}})

LLM Providers

Chat Models

Provider	Config
OpenAI	API key
Anthropic	API key
Groq	API key
Google Gemini	API key
DeepSeek	API key
OpenRouter	API key
AI/ML API	API key
Ollama	Local URL, configurable context window (512 – 131,072 tokens)
LM Studio	Local URL
Custom OpenAI	Base URL + API key + model name

Embedding Models

OpenAI, Ollama, Google Gemini, Xenova Transformers (fully local — no API needed), AI/ML API, LM Studio.

All provider keys are configurable from the Settings UI or config.toml. API keys are never exposed in the frontend.

Personalization & Personas

Personalization

• Location: Optionally bias search results with a configured location.
• About Me: Free-text profile for tone and context (never sent verbatim to external tools).
• Per-message toggles: Enable or disable location/profile injection on each message.

Persona Prompts

Create and manage custom system prompts that shape how the agent responds:

• Built-in templates: Web Searches, Local Documents, Chat Conversations, Scholarly Articles
• One-click copy of templates as starting points for custom personas
• Multiple personas can be active simultaneously
• Stored in the local SQLite database

Installation

Docker (Recommended)

1. Ensure Docker is installed and running.

2. Clone the repository:

   git clone https://github.com/boarder2/Yet-Another-Agentic-Web-Chat.git
   cd Yet-Another-Agentic-Web-Chat

3. Rename sample.config.toml to config.toml and fill in the provider keys you plan to use:

   Key	Required When
   OPENAI	Using OpenAI models
   OLLAMA	Using Ollama (http://host.docker.internal:11434 for Docker)
   GROQ	Using Groq
   OPENROUTER	Using OpenRouter
   ANTHROPIC	Using Anthropic
   GEMINI	Using Google Gemini
   DEEPSEEK	Using DeepSeek
   AIMLAPI	Using AI/ML API
   LM_STUDIO	Using LM Studio

   All keys can also be changed later from the Settings page.

4. Start the stack:

   docker compose up -d

5. Open http://localhost:3000.

Manual Setup

1. Install and configure SearXNG with JSON output enabled.

2. Clone the repo, copy sample.config.toml → config.toml, and fill in your settings.

3. Install dependencies and build:

   npm install
   npm run build
   npm run start

See docs/installation for additional configuration, updating, and tracing setup.

Ollama Connection Errors

OS	Recommended URL
Windows / Mac (Docker)	http://host.docker.internal:11434
Linux (Docker)	http://<host-private-ip>:11434

On Linux, you may also need to set Environment="OLLAMA_HOST=0.0.0.0" in /etc/systemd/system/ollama.service and restart Ollama. See the Ollama FAQ for details.

Using as a Browser Search Engine

1. Open your browser's Search Engines settings.
2. Add a new search engine with URL: http://localhost:3000/?q=%s (replace localhost:3000 with your host and port as needed).
3. YAAWC also exposes an OpenSearch description at /api/opensearch with autocomplete support, so some browsers can discover and add it automatically.

URL queries via ?q= automatically apply your saved model preferences for a seamless search-bar experience.

API

YAAWC exposes a full API for programmatic access:

Endpoint	Method	Description
/api/chat	POST	Streaming chat with tool calls, sources, and live events (SSE)
/api/search	POST	Programmatic search (streaming or non-streaming)
/api/models	GET	List available models (?include_hidden=true for admin view)
/api/config	GET/POST	Read/write server configuration
/api/chats	GET	List all chats
/api/chats/[id]	GET/DELETE	Get or delete a specific chat
/api/suggestions	POST	Generate follow-up suggestions
/api/system-prompts	GET/POST/PUT/DELETE	CRUD for persona prompts
/api/images	POST	Image search
/api/videos	POST	Video search
/api/uploads	POST	File upload
/api/uploads/images	POST/GET	Image upload and serving
/api/tools	GET	List available agent tools
/api/dashboard	GET/POST	Dashboard widget CRUD
/api/respond-now	POST	Interrupt retrieval for immediate response
/api/opensearch	GET	OpenSearch description XML
/api/autocomplete	GET	Search autocomplete (proxied to SearXNG)

For detailed payload schemas, see docs/API/SEARCH.md.

Network & Reverse Proxy

YAAWC runs on Next.js and is accessible on the local network out of the box. For reverse proxy deployments:

1. Set BASE_URL in config.toml under [GENERAL] to your public URL.
2. Forward headers: X-Forwarded-Host, X-Forwarded-Proto, X-Forwarded-Port.

Example Nginx config:

server {
  listen 80;
  server_name yaawc.yourdomain.com;

  location / {
    proxy_pass http://localhost:3000;
    proxy_set_header Host $host;
    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header X-Forwarded-Proto $scheme;
    proxy_set_header X-Forwarded-Host $host;
  }
}

Observability

Built-in support for Langfuse and LangSmith tracing. Configure via environment variables (LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_BASE_URL). See docs/installation/TRACING.md for setup details.

The UI also displays live token usage stats (chat vs. system model), response times, and model names per message.

Contributing

Found a bug? Have an idea? Open an issue or submit a PR. See CONTRIBUTING.md for guidelines.

Acknowledgements

YAAWC is built on the foundation of Perplexica, an open-source AI-powered search engine. We're grateful for their work and encourage you to check out the original project.

A Note on AI Assistance

Yes, a significant portion of this README — and frankly quite a bit of the code in this repo — was written with AI assistance. I know, I know. The irony of using an AI-powered tool to build and document an AI-powered tool is not lost on me. In my defense, I'm a software developer with over 25 years of experience and I review every change, so hopefully the quality bar is somewhat higher than "the AI just vibed it." This is a side project for me to explore agentic architectures and AI tooling.

If you find something that looks suspiciously like a hallucination… well, YOLO, right? Just kidding. Please open an issue and I'll fix it ASAP.