The pplx-embed-v2-late series represents a family of multimodal, late-interaction (ColBERT-style) retriever models built upon the Qwen3.5 architecture. Designed for retrieval tasks involving text, images, and visual documents, these models operate by generating granular, 128-dimensional vector representations for each individual token, which are subsequently compared using the MaxSim similarity metric.

Core Methodology

The models employ a late-interaction paradigm, which facilitates high-fidelity retrieval by preserving token-level information rather than compressing entire documents into a single dense vector. The core similarity mechanism between a query $Q$ and a document $D$ is defined by the MaxSim operator, which computes the maximum cosine similarity of each query vector against the set of document vectors:

$$S(Q, D) = \sum_{q \in Q} \max_{d \in D} \text{cosine}(q, d)$$

This approach allows the model to capture fine-grained contextual alignment between query terms and document segments.

Training and Architecture

The family comprises two primary variants:

• pplx-embed-v2-late-0.6b (340M active parameters): A compact model that is fully fine-tuned to ensure high efficiency.
• pplx-embed-v2-late-9b (7.4B active parameters): A larger, more robust variant where the final eight transformer layers are fully fine-tuned, and the preceding layers and vision encoder are adapted via Low-Rank Adaptation (LoRA).

Both models share a unified embedding space, which enables cross-model compatibility; an index constructed with the 9B model can be queried using the 0.6B model. The models were trained via distillation from a proprietary 18B ColBERT teacher model, utilizing pair and triplet data. The learning process incorporated a token-level LEAF-style objective to align the student model with the teacher’s fine-grained interaction patterns.

Technical Implementation and Constraints

The models utilize bidirectional attention for processing inputs. A critical implementation detail is the placement of special Query/Document tokens; while other systems like PyLate place these at the second position, the pplx-embed-v2-late models expect them at the first position. Furthermore, the models support input batches of either text-only or image-only content, but explicitly do not support mixed text-and-image inputs within a single batch.

Performance

The models achieve high performance on the ViDoRe (v3) benchmark, with the 9B variant reaching 65.2% nDCG@10 for images and 64.7% for markdown documents, significantly outperforming the 0.6B variant (62.3% and 61.2% respectively) while remaining within a highly capable parameter scale for production environments.