Skip to content

Web search #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
b4prog wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Web search #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d6f1a37
[feat] add agentic web research with live activity tracking
b4prog Jul 20, 2026
62259af
[feat] follow relevant links for deeper web research
b4prog Jul 20, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 44 additions & 12 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,18 +13,26 @@ This repository currently provides a clean foundation for Taurus MVP:
- Tauri v2 desktop shell
- Rust backend with typed provider abstraction
- Angular frontend with a chat-oriented UI shell
- Ollama health check, model listing, and streaming chat command flow
- Ollama health check, model listing, streaming chat, and agentic web research flow

## 3. Features

- Local-first architecture (no telemetry and no remote SaaS dependencies by default)
- Local-first architecture with no telemetry; external web requests happen only when the model uses a web tool
- Provider abstraction (`ChatProvider`) to support additional providers later
- Ollama integration:
- availability check
- automatic startup health check
- automatic model listing when Ollama is reachable
- model listing
- streaming chat request/response (real-time token streaming)
- Agentic web research:
- `search_web` searches the public web through Bing's structured RSS results
- `fetch_web_page` extracts readable text and labeled links from a public HTTP or HTTPS page
- the two highest-ranked search results are fetched automatically so research uses page content rather than snippets alone
- the planner can selectively follow relevant links from hubs, indexes, directories, and overview pages to reach more specific documents
- bounded planning rounds and tool calls fall through to a best-effort answer instead of interrupting the workflow
- private/local network destinations and non-text downloads are rejected
- Live agent activity in the chat, with one expandable row per planning, search, page-read, source-coverage, or writing step
- Typed command contracts between Angular and Rust
- Basic chat workspace UI:
- conversation sidebar placeholder
Expand All @@ -45,7 +53,9 @@ Backend (Rust):
- Tauri commands are thin and validate input
- Provider-independent types live in `providers/mod.rs`
- Ollama integration stays in `providers/ollama.rs`
- Shared application state stores provider instances
- `agent.rs` coordinates model/tool turns without coupling the UI to Ollama
- Web tool execution and network safety rules live under `tools/`
- Shared application state stores provider and tool instances

## 5. Repository Structure

Expand All @@ -72,15 +82,19 @@ Backend (Rust):
├── tauri.conf.json
└── src/
├── app_state.rs
├── agent.rs
├── error.rs
├── commands/
│ ├── chat.rs
│ ├── health.rs
│ ├── mod.rs
│ └── models.rs
└── providers/
├── providers/
├── mod.rs
└── ollama.rs
└── tools/
├── mod.rs
└── web.rs
```

## 6. Requirements
Expand Down Expand Up @@ -250,7 +264,15 @@ npm install

This will generate and update `package-lock.json`.

## 14. Running Taurus in Development Mode
## 14. Using Web Research

Choose an Ollama model that supports tool calling. When a prompt needs current or external information, the model can search the web. Taurus then fetches the two highest-ranked result pages so the model can use page content in its final response. If a fetched result is a hub, index, directory, feed, listing, or overview, Taurus exposes its labeled links to the planner so the planner can open only the documents relevant to the request. For synthesis, comparison, explanation, and evaluation tasks, headlines and short summaries on a hub are treated as discovery material; the planner is instructed to fetch a small representative set of the linked documents before answering. This behavior is generic and is not tied to news or any particular website structure.

Each action appears above the assistant response as a one-line status row. Select a row to expand or collapse it. Search details include the exact query, result titles, URLs, and snippets. Page-read details include the fetched text and a bounded list of discovered links. The source-coverage step reports how many search result sets and pages were collected, whether more evidence is needed, and the exact next action.

Web research is opt-in at the prompt level: the model decides when it is needed. Search queries are sent to Bing, and the highest-ranked result pages are requested directly from their hosts. Do not include secrets in a prompt that asks for web research.

## 15. Running Taurus in Development Mode

Run the desktop app (Angular + Tauri):

Expand All @@ -265,7 +287,7 @@ npm run start
npm run build
```

## 15. Building the Application
## 16. Building the Application

Production desktop build:

Expand All @@ -279,7 +301,7 @@ Frontend-only production build:
npm run build
```

## 16. Running Tests
## 17. Running Tests

Frontend unit tests:

Expand All @@ -295,7 +317,7 @@ npm run rust:test

If frontend tests fail due missing browser runtime, install Chrome/Chromium and retry.

## 17. Formatting the Code
## 18. Formatting the Code

Format and check formatting:

Expand All @@ -305,7 +327,7 @@ npm run format:check
npm run rust:fmt
```

## 18. Linting the Code
## 19. Linting the Code

Frontend lint:

Expand All @@ -319,22 +341,29 @@ Rust lint:
npm run rust:clippy
```

## 19. Type Checking
## 20. Type Checking

Run TypeScript type check:

```bash
npm run typecheck
```

## 20. Troubleshooting
## 21. Troubleshooting

Common issues:

- Ollama unavailable:
- Make sure Ollama is running.
- Verify `TAURUS_OLLAMA_BASE_URL`.
- Confirm `http://localhost:11434/api/tags` is reachable.
- Agent planning fails with a tool-calling error:
- Select an Ollama model with tool-calling support.
- Confirm the model can use tools through Ollama's `/api/chat` endpoint.
- Web search or page fetch fails:
- Confirm the machine has internet access.
- Bing or the target site may reject automated requests or apply rate limits.
- Localhost, private IP ranges, oversized responses, and binary content are blocked intentionally.
- Linux build errors about WebKit/GTK:
- Recheck distro package names from Tauri prerequisites docs.
- `npm run test` fails in headless environments:
Expand All @@ -352,9 +381,12 @@ Common issues:
npx -y node@20 ./node_modules/@angular/cli/bin/ng build
```

## 21. Security Notes
## 22. Security Notes

- Taurus is local-first and does not include telemetry by default.
- Web research sends the model's search query to Bing and automatically fetches up to two highest-ranked public result pages per search. The planner may selectively request additional public pages linked by fetched results, subject to the global planning-round and tool-call limits.
- The fetch tool blocks loopback, private, link-local, and other special-use IP ranges, including across redirects.
- Web responses are size-limited and page text is treated as untrusted reference material in the agent prompt. The expandable workflow retains the fetched text, while the model receives a bounded excerpt to leave enough context for its answer.
Comment on lines +384 to +389

@coderabbitai coderabbitai Bot Jul 20, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Renumber the Roadmap section.

These changes make Security Notes section 22, but the later Roadmap heading remains section 22 at Line 396. Rename the Roadmap heading to ## 23. Roadmap.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@README.md` around lines 384 - 389, Update the later Roadmap heading in
README.md from section 22 to section 23, preserving the existing Roadmap
content.

- Tauri command surface is intentionally narrow:
- no arbitrary shell execution
- no arbitrary filesystem commands
Expand Down
2 changes: 1 addition & 1 deletion angular.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@
},
{
"type": "anyComponentStyle",
"maximumWarning": "4kB",
"maximumWarning": "6kB",
"maximumError": "8kB"
}
],
Expand Down
Loading
Loading