Skip to content

Releases

v2026.09.5 (0.1.6)

Choose a tag to compare

@vectorequa vectorequa released this 18 Sep 15:58

min_image_dim support + 4K resolution for qwen3.6-35b

min_image_dim — per-provider minimum image dimension

Add min_image_dim to ProviderCapabilities and VisionConfig. Images smaller than the minimum are automatically upscaled using sharp fit:outside before sending to the model.

4K resolution support

Updated qwen3.6-35b config based on model's preprocessor_config.json:

Parameter Before After Source
max_image_dim 2048 4096 size.longest_edge = 16777216 = 4096²
min_image_dim (none) 256 size.shortest_edge = 65536 = 256²

Files changed

  • src/config.ts: ProviderCapabilities + VisionConfig add min_image_dim, resolveProviderImageDim resolves it per-provider
  • src/image-loader.ts: preprocess() upscales images below min_image_dim using sharp fit:outside, withoutReduction:true
  • src/tools.ts: computeImageOverrides passes minImageDim through

Test

  • 35 existing tests pass
  • Verified: 16x16 image upscaled to 32x32 with minImageDim=32

v2026.09.4 (0.1.5)

Choose a tag to compare

@vectorequa vectorequa released this 18 Sep 11:09

extra_body — per-provider vendor-specific parameter passthrough

Add extra_body to any provider config to pass vendor-specific fields directly into the LLM request body. Standard fields (model, messages, max_tokens) always override.

Primary use case: disable thinking on reasoning models (qwen3.6-35b vLLM) for 10-50x faster OCR/vision tasks.

Example config

"qwen": {
    "extra_body": {"chat_template_kwargs": {"enable_thinking": false}}
}

Measured impact (qwen3.6-35b, same OCR task, max_tokens=50)

thinking OFF (extra_body) thinking ON (before)
Latency 314ms 676ms
Content "Yellow" "" (empty!)
Tokens 2 50 (all reasoning)
Finish stop length (truncated)

Safety

  • Providers without extra_body → no extra params sent → zero impact
  • 400 graceful degradation: if backend rejects extra_body, auto-retry without it

Future-proof

Handles any vendor's thinking control via config — no code changes for new models:

  • vLLM/Qwen: {"chat_template_kwargs": {"enable_thinking": false}}
  • OpenAI: {"reasoning": {"effort": "low"}}
  • DeepSeek: {"reasoning_effort": "low"}
  • Gemini: {"thinkingLevel": "low"}

v2026.09.3 (0.1.4)

Choose a tag to compare

@vectorequa vectorequa released this 18 Sep 04:00

Changes

feat: extract_text custom prompt + simplified OCR default (494cc45)

  • extract_text: add optional prompt param for custom OCR instructions
  • Simplify DEFAULT_OCR_PROMPT (auto-detect instead of hardcoded language list)

feat: SSRF protection, tool annotations (c6f6c36)

  • SSRF: validateImageUrl() blocks private IPs (10.x / 172.16-31.x / 192.168.x / 169.254.x / fc00::/7 / fe80::/10); loopback allowed for local stdio process
  • Tool annotations: all 4 tools declare readOnlyHint: true + openWorldHint
  • Tool title: Analyze Image / Extract Text (OCR) / Extract Structured Data / Ping

fix: MCP spec compliance (80e2b41)

  • errorResponse returns isError: true with plain text (not JSON)
  • extract_structured parse failure throws instead of fake success
  • default_provider routing respects fallback exclusion
  • EXIF auto-rotation via sharp.rotate()
  • Image dimension logging in all handlers

v2026.09.0 (SemVer 0.1.1)

Choose a tag to compare

@vectorequa vectorequa released this 16 Sep 00:10

Initial release after orphan reset.

Features

  • 6 MCP tools: analyze_image, extract_text, extract_structured, compare_images, analyze_screenshot, ping
  • 4 MCP prompts: ocr, describe, compare, ui-tree
  • Any OpenAI-compatible vision API (Qwen-VL, GPT-4o, GLM-4V, Gemini, local vLLM, etc.)
  • Multiple input formats: file path, base64, URL
  • Automatic image preprocessing via sharp (optional)
  • bbox coordinate scaling for screenshot-to-accessibility-tree
  • Structured logging with auto-rotation
  • Retry with backoff, JSON-mode fallback
  • Per-provider capabilities and enable/disable control
  • Config hot-reload without restart
  • Zero Python dependency — pure TypeScript/Node.js