Result from a retrieval operation.

Raw capture result for one tile. Decoded later during verification.

Raw capture results for all tiles of one article.

Learnable logit scale (from contrastors/OpenCLIP). Contrastors uses a learnable log-scale initialized to ln(1/0.07) ≈ 2.66. The parameter is

Direct websocket CDP connection.

Raised when text LLM calls fail after retries.

Raised when VLM API calls fail after retries.

Dataset of (query, positive_path, [negative_paths...]) from JSONL. Supports two formats: - Simple: {"query": "...", "chunk_path": "..."}

Retrieve text chunks from a text search API (wiki-screenshot text_search_api.py). The API accepts: POST /search {"queries": [{"te

Use Jina Reader API to fetch clean markdown text from URL. Jina Reader (r.jina.ai) converts any URL to LLM-friendly markdown text.

Simplified async LLM client for Gemini using Vertex AI.

Retrieve tiles from a local search API (e.g. localhost:30888/search). The API accepts batch queries: {"queries": [{"text": "..."}, ...],

Raised when a referenced image path no longer exists on disk.

WorldVQA without retrieval: query + image from dataset only. WorldVQA images are embedded in the HuggingFace dataset (PIL or base64).

Sequential tile capture. One screenshot at a time per page. Args: chrome_path: Chrome binary path. n_workers: Number of Chrome pr

Metadata for a single chunk image (1024px strip of a tile).

Retrieve similar screenshots using ColQwen2 LEANN multi-vector retrieval.

Use ds-serve API for external text augmentation. Calls ds-serve search API to retrieve relevant text passages for the query.

Exponential Moving Average of model parameters for better generalization.

EVQA without retrieval: query + iNaturalist image only, no Wikipedia tiles. Used to test VLM's ability to answer from the species image alone.

Text-retrieve → DOM lookup: retrieve text chunks, then find their HTML context. Wraps TextAPIRetriever. For each retrieved text chunk: 1. Fet

Merge image (LocalAPIRetriever) and text (TextAPIRetriever) hits by raw score. Both underlying retrievers embed with Qwen3-VL-Embedding-2B agains

Manages multiple kiwix-serve processes for high-concurrency serving. Benchmark results (80 concurrent requests, 2000 articles): 1x --thre

Ground-truth tiled retriever using pre-rendered Wikipedia tiles from local kiwix. For each example, looks up the Wikipedia URL in the local kiwix

Yields batches with exact per-source counts from a ConcatDataset. Each source occupies a contiguous index range [start, start+length). Every

No retrieval - returns empty result, LLM answers from its own knowledge.

Wraps an image retriever; OCRs retrieved tiles and returns text. Ablation A pipeline: image retrieve -> OCR -> text to reader. Talks to an Op

Pool of persistent one-GPU workers with dynamic work distribution. Workers share a single work queue. Fast GPUs pull more work items, elimina

Renders text queries as small, clear PNG images. Images are cached on disk so each query is only rendered once.

Playwright-managed CDP connection.

Dataset of (query, chunk_path) pairs from JSONL.

Renders query text + query image together into a single card image. Reuses PixelQueryRenderer for text rendering. Layout: image on top, text belo

Save/restore RNG state for GradCache replay (from contrastors/rand_state.py).

Wraps a text retriever; renders each chunk as an image. Ablation B pipeline: text retrieve -> render as Wikipedia-style image -> VLM reader.

Coordinates shard distribution across machines via S3 claim files. Each machine runs a claim loop: while True: shard = coordi

Use screenshot that was prepared in data layer. Expects screenshot to be captured beforehand. This retriever just loads and encodes the exist

Dataset of (query, positive_text, [negative_texts...]) from text-qa-pair JSONL. Each row has: query, passage (positive), neg_passages (list of ha

Use text content fetched from URL. Can use pre-cached text or fetch on demand.

Retrieve text using LEANN vector search. Uses LEANN's integrated embedding + indexing system for text retrieval. Supports various embedding m

Metadata for a single tile image (or a chunk of one).

Retrieve similar image tiles using ColQwen2 LEANN multi-vector retrieval. Splits dataset screenshots into fixed-size tiles, embeds each tile with

Retrieves context by searching through image tiles using Qwen3-VL-Embedding. Uses single vector embeddings (2048 dim) with cosine similarity for

Use tiled screenshot from ground truth URL. Captures screenshot for the example's URL, splits it into tiles, and returns tiles. This is groun

Retrieve similar image tiles using vector similarity search. Splits dataset screenshots into fixed-size tiles, embeds each tile, and retrieve

Chrome process with two tabs, each independently controllable.

Retrieve similar screenshots using vector similarity search. Uses Jina API for embedding and retrieval across dataset screenshots only.

Use Wikipedia API to fetch clean article text. Extracts Wikipedia page title from URL and fetches content via API. Much cleaner and faster th

Minimal CDP connection with a receive loop.

Base class for retrieval strategies.

Benchmark harness. Measures throughput/latency and verifies correctness. Usage: bench = Bench(zim_path="...", chrome_path="...", output_d

Single-vector bi-encoder with last-token pooling + L2 normalization.

Parallel tile capture via directClip. For each viewport-sized chunk of the page: 1. Scroll to chunk position, wait 2 rAF 2. Fire N parall

Capture tiles dynamically — no upfront height measurement. 1. Navigate, wait for load event only (no fonts.ready) 2. Capture tile at y=0 imme

Single full-page screenshot, crop tiles in Python. 1. Set viewport to 1080px, navigate, measure scrollHeight (true height) 2. Inject CSS to l

Multi-tab igpu: fewer processes, more tabs per process.

Capture tiles using clip.y offset without scrolling. Sets viewport once, then captures each tile region via clip parameter with captureBeyond

One-shot: fresh headless_shell process per article. n_workers controls how many articles are captured concurrently (each in its own short-liv

2-tab overlapping nav+capture with semaphore-limited captures.

Fire all tile screenshots simultaneously per page. Same setup as CDPSequentialStrategy, but within one page, all tiles are sent at once witho

Fonts-only nav + per-tile viewport image wait. Fast nav (~70ms), images loaded on-demand per tile.

Semaphore-limited concurrent captures. n_workers Chrome processes, but only capture_limit capture simultaneously. The rest are free to naviga

Two-tab ping-pong: overlap nav (I/O) with capture (mem BW). Each Chrome process has 2 pages. While one captures, the other loads.

Interface for capture strategies.

Abstract connection to one Chrome process.

Data source backed by a local Kiwix ZIM file served via kiwix-serve. Both article HTML and embedded images are served from the ZIM archive, e