pplx-embed-v2-context-9b

Name: pplx-embed-v2-context-9b
Author: perplexity-ai

by perplexity-ai

Contextual embeddings for text chunks and page images, sharing a 1024-d space with the 0.6B model

13likes

8.4Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

perplexity-ai/pplx-embed-v2-context-9b

Feature URI

Deploy pplx-embed-v2-context-9b

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

Read the customer-hosted guide Talk to us about single-tenant

Overview

pplx-embed-v2-context-9b is the full release of Perplexity's contextual embedding model, published on 28 October 2026 under the MIT license. It embeds each chunk of a document with the rest of the document in view, and it now embeds images too: a page image is split into horizontal bands, one vector per band, so a slide or a scanned page can be searched by its parts.

Its first 1024 dimensions share a space with pplx-embed-v2-context-0.6b, so you can embed documents with one model and queries with the other. The technical report (arXiv 2610.38854) says the 9B model sets the state of the art on its two new chunk benchmarks, Context-bench and Q2D-Chunk, and on ConTEB, and leads dense retrievers on ViDoRe V3 for visual documents. The scores are Perplexity's own.

Architecture

A vision-language backbone of about 8.4B parameters with custom code (trust_remote_code, transformers 5.4 or later), fine-tuned from pplx-embed-v2-late-9b. Training distils graded chunk-level relevance from a query-aware context-compression model, so it needs no gold-chunk labels. Text chunks of one document are encoded in one pass; images are encoded in one pass with their visual tokens divided into whole-row bands. Matryoshka losses at 1024 and 2048 dimensions; output is unnormalized int8.

Mixpeek SDK Integration

# No Mixpeek extractor takes a Hugging Face model id, so this model runs on your side
# and its chunk vectors are upserted into a bring-your-own-vectors collection.
# Chunks of one document must not overlap.
from transformers import AutoModel
import requests

model = AutoModel.from_pretrained("perplexity-ai/pplx-embed-v2-context-9b", trust_remote_code=True).to("cuda")
chunks = ["Q3 revenue grew 14 percent.", "Most of the growth came from EMEA.", "Churn fell to 2 percent."]
vectors = model.encode([chunks], normalize_embeddings=True)[0]  # one 2048-d vector per chunk, in context

requests.post(
    "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
    headers={"Authorization": "Bearer API_KEY"},
    json={"collection_id": "col_your_collection",
          "documents": [{"document_id": f"report-q3-{i}", "vectors": {"text-embedding": v.tolist()},
                         "payload": {"text": c, "chunk": i}} for i, (c, v) in enumerate(zip(chunks, vectors))]},
)

Capabilities

Contextual chunk embeddings for text, encoded together per document
Image embeddings per horizontal band of a page (encode_images, 8 bands by default)
Separate query encoding (encode_queries) from document encoding (encode)
2048 dimensions, or 1024 by truncation, shared with the 0.6B model
Multilingual; MIT license

Use Cases on Mixpeek

RAG over long reports and contracts where a chunk alone is ambiguous

Searching slides, scans and PDF pages as images, scored by their best band

Embedding a corpus with the 9B model and serving queries with the 0.6B model at 1024 dimensions

Performance

Input SizeDocuments as lists of non-overlapping text chunks, or page images split into bands; queries encoded separately

Embedding Dim2048 (1024 by Matryoshka truncation, shared with the 0.6B model)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

Benchmark claims are from Perplexity's report and we have not measured them. Output is unnormalized int8; compare with cosine similarity or normalize first. Vectors from the earlier preview should not be mixed with this release's. Needs a GPU for practical throughput.

Common Pipeline Companions

perplexity-ai/pplx-embed-v2-context-0.6b

The small model with the same 1024-d space, for cheap query encoding

perplexity-ai/pplx-embed-v2-context-9b-preview

The September preview this release replaces

Frequently Asked Questions

What changed from pplx-embed-v2-context-9b-preview?

This is the final release. It adds image embeddings through encode_images, which splits a page into horizontal bands with one vector each, and its 1024-d prefix shares a space with the 0.6B model. Re-embed anything you indexed with the preview, since the two releases' vectors are not interchangeable.

How do I search page images with it?

Call encode_images on the page with chunks=8 (the default) to get one 2048-d vector per band, encode the query with encode_queries, and score the page by its highest band similarity. Store each band as its own vector so a standard ANN index can find it.

Can I index with the 9B model and query with the 0.6B model?

Yes, at 1024 dimensions. Take the first 1024 values of each unnormalized vector from either model and normalize afterwards. Those are the only shared dimensions; 2048-d vectors are 9B only.

How do I encode queries with pplx-embed-v2-context-9b?

Use encode_queries for queries and encode (or encode_images) for documents. The model was trained with separate query and document prefixes, and the card warns that encoding queries with encode silently lowers retrieval quality.