# Rerank first-stage results before they reach the model

> Rerank first-stage results before they reach the model: 2 companies fully do it, 3 with a limit, 14 carry the tag. Every listing in Embeddings & rerankers with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/embeddings.rerank-candidates/
- Markdown: https://www.anchorterminal.com/niches/embeddings.rerank-candidates/index.md (~1,600 tokens)
- Slim: https://www.anchorterminal.com/niches/embeddings.rerank-candidates/index.min.md (~280 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/embeddings.rerank-candidates/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 3 | 14 | 8 | 7 |

## What counts as done

One call takes the query and all candidates and returns each candidate's index with a relevance score, sorted or sortable, with a documented per-document token limit and truncation behaviour and a published billing unit.

Measured, for our own tests: 50 candidates reranked within 1 second at p95, and nDCG@10 above the first-stage order on the labelled set.

How agents ask for it: rerank these results; which of these chunks answers the question; improve the precision of my RAG; sort these 50 passages by relevance

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Jina Embeddings and Reranker (Jina AI GmbH)**: POST /v1/rerank returns candidates sorted by relevance_score with their index, top_n is optional, max_doc_length defaults to 8,192 on v3.5 in a 131,072-token window, and the catalogue bills $0.05 per million tokens. p95 and nDCG unmeasured.
- **Voyage AI embeddings and rerankers (Voyage AI Innovations, Inc.)**: One rerank call takes a query and up to 1,000 documents, with top_k, per-model query and query-plus-document token limits, truncation true or false documented, and per-token prices. p95 and nDCG unmeasured.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Cohere Embed and Rerank (Cohere Inc.)**: One call returns index and relevance_score with top_n; max_tokens_per_doc defaults to 4,096 and truncates silently, and the billing unit is unclear, with billed_units carrying both tokens and search_units. Accounts report large quality gains.
- **NVIDIA NeMo Retriever Embedding and Reranking NIMs (NVIDIA Corporation)**: /v1/ranking scores up to 512 passages in one call and returns index and logit sorted descending, with truncate NONE or END documented, but there is no top-n field and billing is per GPU, not per call.
- **ZeroEntropy zerank and zembed (ZeroEntropy, Inc.)**: The hosted rerank endpoint is gone. zerank-2 is Apache 2.0 with a Sentence Transformers route and 32,768-token context, but a May 2026 breaking change made predict() return raw logits, and no per-call limits or billing remain.

- not this job: Amazon Nova Multimodal Embeddings (Amazon Web Services, Inc.): An embedding model only; no rerank call in the Nova guide or schema.
- not this job: EmbeddingGemma 2 (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)): An embedding model; no reranking head or scoring mode.
- not this job: EmbeddingGemma 300M (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)): An embedding model; no reranking mode.
- not this job: Gemini Embedding (Google LLC): No rerank endpoint in the embeddings guide or model page.
- not this job: Granite Embedding Small English R2 (International Business Machines Corporation): A bi-encoder only; no reranking mode in this listing.
- not this job: jina-embeddings-v5-omni-small (Jina AI GmbH, a subsidiary of Elastic N.V. (per the jina-embeddings listing's 2 October 2026 check)): An embedding model; no reranking head.
- not this job: Mistral Embed and Codestral Embed (Mistral AI (RCS Paris 952 418 325)): No rerank endpoint in the OpenAPI document.
- not this job: Nomic Embed (Nomic, Inc.): No rerank endpoint in the API reference.
- not this job: OpenAI embeddings (OpenAI OpCo, LLC): No rerank endpoint.
- not this job: Qwen3-Embedding-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check)): An embedding model; the Qwen3-Reranker models are separate repositories, not this listing.
- not this job: Qwen3-Embedding-4B (Alibaba Cloud (named in the copyright line of the 8B sibling's Apache 2.0 licence; this repository has no LICENSE file)): An embedding model; the Qwen3-Reranker models are separate repositories, not this listing.
- not this job: Qwen3-Embedding-8B (Alibaba Cloud (named in the licence's copyright line)): An embedding model; the Qwen3-Reranker models are separate repositories, not this listing.
- not this job: Qwen3-VL-Embedding-2B (Alibaba Cloud (RDAP names Alibaba Cloud (Singapore) Private Limited as the registrant of qwen.ai)): An embedding model; the Qwen3-VL-Reranker models are separate repositories, not this listing.
- not this job: Qwen3-VL-Embedding-8B (Alibaba Group (Tongyi Lab, named in the copyright line of the GitHub repository's Apache 2.0 licence)): An embedding model; the Qwen3-VL-Reranker models are separate repositories, not this listing.

## What stops agents finishing

- default per-document token cap truncates long chunks (Cohere max_tokens_per_doc 4,096)
- no top-n field, so the agent must sort and cut (NeMo Retriever /v1/ranking)
- billing unit unclear (searches against tokens)
- the reranker only reorders what the candidate stage found

## Around it

Before: vector-search.hybrid-keyword, embeddings.index-docs-for-rag. After: company-knowledge.cited-answer.

## Where this comes from

Weight 8 of 100 in Embeddings & rerankers: 7 of 62 accounts (91 account-use-case tags, accounts counted in each use case they fit, scaled to 100 and rounded by hand); every account that tried a reranker called it a large gain; fewer accounts than its importance in threads suggests, so held at the count

As data: https://www.anchorterminal.com/api/v1/use-cases/embeddings.rerank-candidates.json, or the MCP tool get_use_case with id embeddings.rerank-candidates. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
