The interesting release this week wasn't a frontier model. It was the embedding layer.
On October 7, Perplexity released two MIT-licensed embedding models, pplx-embed-v2-late, in a 0.6B edge size and a 9B production size. Both use ColBERT-style late interaction and retrieve across text, images, and rendered PDF pages without an OCR preprocessing step. Google also released EmbeddingGemma 2 under Apache 2.0, a multimodal embedding model aimed at local search and retrieval that maps text, code, images, video, and audio into a shared 768-dimensional space.
Taken together, two things that were previously either proprietary or research-grade are now self-hostable with permissive licenses: late-interaction scoring, and page-image retrieval that skips text extraction entirely.
What actually changes in the pipeline
If you index rendered pages instead of extracted text, the most failure-prone stage of most knowledge bases disappears. OCR and layout parsing is where multi-column reports, merged table cells, footnotes, and scanned appendices quietly turn into garbage that no reranker can rescue. It also makes a lot of chunking strategy moot — which lands next to a placebo-controlled study circulating this week that challenges the assumption that structure-aware chunking is doing the work people credit it with.
The cost is real and it's at the index. Late interaction keeps a vector per token or patch rather than one per chunk, so your vector count grows by orders of magnitude for the same corpus, and scoring becomes MaxSim over those vectors rather than a single dot product. You need pooling, quantization, or a two-stage candidate generation step before this is affordable at enterprise scale. Budget for storage and latency work, not a drop-in swap.
The second cost is subtler: no text layer means no BM25 side of your hybrid retrieval, no keyword filters, no regex PII scan before indexing, and citations that point at a page image rather than a span. For regulated corpora that's a governance problem, not a tuning problem. The pragmatic shape is both — extract text for filtering, lexical matching, and provenance, embed the rendered page for recall.
Who should care, and when
Teams with PDF-heavy corpora in finance, legal, and engineering documentation get the biggest lift, because that's where extraction loses the most information. Teams already on a managed embedding API now have a permissive alternative they can run inside their own boundary, which matters more for data residency than for quality. The 0.6B model and EmbeddingGemma 2 put usable retrieval on-device, which is the first credible path to local personal knowledge indexes that don't ship documents anywhere.
Context pricing moved in the same direction: Claude Haiku 5.5, released October 7, has a 1M-token context window at $0.10 per million input tokens. Cheap long context makes LLM reranking over a wide candidate set viable, which partly compensates for a lossier first-stage retriever. It does not replace retrieval. Stuffing a million tokens is still a worse answer than fetching the right fifty.