Understand an image and return a detailed text description.

Use this when you need to "see" an image — DeepSeek-V4 cannot read images directly. Accepts an image as a local file path, an HTTP(S) URL, a data URI, or a base64 string. Optionally pass question to focus the analysis (e.g. "What error is shown in this screenshot?", "Read the chart values."). Returns structured text you can reason over. Requires a configured vision provider (VISION_*); degrades to metadata + OCR without one.

Extract readable text from an image (OCR).

Accepts a file path, URL, data URI, or base64 string. Uses local Tesseract when installed; otherwise falls back to the configured vision provider. lang is a Tesseract language code (e.g. 'eng', 'chi_sim', 'eng+chi_sim').

Search the web and return ranked results (title, URL, snippet).

Use this for live, up-to-date information. Returns up to max_results hits as a numbered markdown list. Follow up with fetch_page to read any result in full. Works without any API key (DuckDuckGo).

Fetch a web page and return clean, readable Markdown.

Use this to read the full content of a URL (e.g. a search result). Strips navigation/ads and extracts the main article text. Output is capped to max_chars characters. Works without any API key.

Run a multi-step deep-research investigation and return a cited report.

Pipeline: decompose query into search queries -> search the web -> fetch & read pages -> (between rounds) refine queries from findings -> extract the most relevant passages -> synthesise a structured markdown report with inline [n] citations and a numbered source list. Uses DeepSeek-V4 (flash for plan/refine/rerank, pro for synthesis).

Tunables (0 = use defaults from config): breadth sub-queries per round (default 3), depth rounds (default 2), max_sources pages read (default 8). Higher = more thorough but slower/costlier. Requires DEEPSEEK_API_KEY; search/fetch are keyless (DuckDuckGo).