DuckDuckGo Reworked for LM Studio

Reliable DuckDuckGo web and image search tools for LM Studio with stronger anti-block resilience, better output consistency for LLM tool chaining, and cleaner diagnostics.

What was broken before

The previous implementation had several practical issues in real usage:

Web search sometimes returned only https://duckduckgo.com/ instead of actual results.
DDG redirect links were not decoded correctly, so target URLs were lost.
Internal DuckDuckGo links polluted the final output and looked like "valid" search hits.
Label extraction was noisy because HTML text was not normalized robustly.
Image search frequently failed with Forbidden or Unable to extract vqd token with no useful fallback path.
Image filtering was too strict (URL extension based), dropping valid CDN images without .jpg/.png suffixes.
Download stage could report confusing outcomes (including false-positive success patterns in older behavior).
Tool-output shape encouraged weak orchestration in some models (extra unnecessary calls to downstream image tools).

What is fixed in this reworked version

Web Search quality improvements

Uses https://html.duckduckgo.com/html/ for more stable HTML result pages.
Replaces fragile URL regex extraction with anchor-level parsing (<a href=...>).
Decodes DuckDuckGo redirect links via uddg into real destination URLs.
Filters out internal DuckDuckGo service/home links from results.
Cleans HTML labels and normalizes whitespace for readable output.

Image Search reliability improvements

Added resilient retry/backoff for DDG requests (vqd bootstrap and i.js endpoint).
Added improved request header profiles for document/json/image fetches.
Added in-memory cache for image query results and vqd tokens.
Added content-type validation (image/*) during downloads instead of extension-only acceptance.
Added explicit structured block diagnostics:
- error: blocked_by_waf
- stage: vqd | i.js
- try_web_fallback: true

Multi-source fallback (no API keys)

When DDG image endpoint is blocked or unstable, Image Search now automatically runs a fallback chain:

DuckDuckGo web results (html.duckduckgo.com)
Extract image candidates from top pages (src, data-src, data-original, srcset, markdown image links)
Download and validate candidates (content-type: image/*)
Return renderable local markdown images

Fallback responses include source: "web_fallback" when that path was used.

Output consistency for LLM tool use

Image Search returns renderable markdown images in a consistent shape:
- images
- count
- hint (that images are already downloaded/renderable)
This reduces redundant tool hops and improves assistant decision quality in multi-tool flows.

Better result reliability (general)

Preserves deduplication, pagination (page, pageSize), and safeSearch behavior.

Installation

Install from LM Studio Hub:

vadimfedenko/duck-duck-go-reworked

Then click Run in LM Studio.

Configuration

Search Results Per Page: max number of search results (0 = auto).
Safe Search: off, moderate, strict (auto lets assistant decide).

Usage

With the plugin enabled, the assistant can:

run web search (Web Search)
run image search (Image Search)
paginate results when needed

Works especially well with the Analyze Images and improved Visit Website plugin:

LLM can find images via this plugin, then analyze the content with analyze.

Changelog (reworked release)

Fixed homepage-only result regression.
Added DDG redirect target decoding.
Added internal-link filtering.
Improved title/label extraction cleanliness.
Added robust image fetch retries and anti-block diagnostics.
Added automatic web-fallback image pipeline when DDG image endpoint is blocked.
Added output consistency hints for better model tool orchestration.

duck-duck-go-reworked