pplx-embed-v2-context-9b-preview

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

by perplexity-ai

Contextual chunk embeddings: each chunk is encoded with the rest of its document, at 2048 or 1024 dimensions

52likes

8.4Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

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

Feature URI

Deploy pplx-embed-v2-context-9b-preview

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-preview embeds the chunks of a document together, so each chunk's vector reflects the text around it. That helps retrieval when a chunk on its own is ambiguous, which is most chunks in long reports and contracts. Perplexity released it on 25 September 2026 under the MIT license as a preview.

It outputs 2048 dimensions, or 1024 by truncation, as unnormalized int8 values, and encodes queries with a separate method from documents.

The card reports no benchmark scores, and as a preview its weights and embeddings may change without backward compatibility.

Architecture

A roughly 8.4B-parameter transformer with custom code (trust_remote_code). A document is passed as a list of chunks and encoded in one pass, and each chunk's embedding is the mean of its token states, so it reflects the whole document. Queries use fixed query prefixes through encode_queries. Trained with Matryoshka losses at 1024 and 2048 dimensions and quantized to 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.
from transformers import AutoModel
import requests

model = AutoModel.from_pretrained("perplexity-ai/pplx-embed-v2-context-9b-preview", 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 embeddings for document chunks, encoded together per document
Separate query encoding (encode_queries) from document encoding (encode)
2048 dimensions, or 1024 by truncation (Matryoshka); int8 output
Multilingual; MIT license; preview release

Use Cases on Mixpeek

Retrieval over long documents where a chunk alone is ambiguous ("it", "the company", "this quarter")

RAG over reports, contracts and transcripts split into chunks

Testing contextual embeddings against a plain chunk embedding on your own corpus

Performance

Input SizeDocuments as lists of text chunks, encoded together; queries encoded separately

Embedding Dim2048 (1024 by Matryoshka truncation)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

The card reports no benchmark scores for this preview. Output is unnormalized int8; compare with cosine similarity or normalize first. Preview embeddings should not be mixed with a later version's. Needs a GPU for practical throughput. We have not measured it.

Common Pipeline Companions

perplexity-ai/pplx-embed-context-v1-4b

The earlier 4B contextual model, with published scores

perplexity-ai/pplx-embed-v1-0.6b

A small non-contextual Perplexity embedder for cheaper indexing

Frequently Asked Questions

What is a contextual embedding model?

It embeds each chunk of a document with the other chunks in view, so a chunk that says "revenue grew 14 percent" carries which company and which quarter from elsewhere in the document. A plain embedding model sees each chunk alone.

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

Use encode_queries for queries and encode for document chunks. The model was trained with different prefixes for each, and the card warns that encoding queries with encode gives worse results.

How many dimensions does it output?

2048, or 1024 if you take the first 1024 values and normalize afterwards. Those are the two sizes it was trained for; other truncations were not trained.

Can I use it in production?

It is a preview. Perplexity says weights, embeddings and the interface may change without backward compatibility, so do not mix its vectors with a later version's, and plan to re-embed when the final model ships.