# Embed a large corpus in bulk and keep the index small

> Embed a large corpus in bulk and keep the index small: 2 companies fully do it, 16 with a limit, 35 carry the tag. Every listing in Embeddings & rerankers with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/embeddings.bulk-backfill/
- Markdown: https://www.anchorterminal.com/niches/embeddings.bulk-backfill/index.md (~2,150 tokens)
- Slim: https://www.anchorterminal.com/niches/embeddings.bulk-backfill/index.min.md (~630 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/embeddings.bulk-backfill/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 | 16 | 35 | 12 | 11 |

## What counts as done

A batch or async route takes the corpus and returns precisely one vector per input in input order, its price per million tokens (and any batch discount) is published, and vectors can be shortened or returned as int8 or binary through a documented parameter that the chosen route honours.

Measured, for our own tests: 1M tokens embedded without an unhandled 429, and wall-clock time for that job.

How agents ask for it: embed all 10 million records; backfill embeddings for the whole archive; cut our vector storage bill; give me int8 or binary vectors

## 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/batch/embeddings takes OpenAI-style JSONL keyed by custom_id; binary, ubinary and dimensions are request fields; the model catalogue prices v5-small at $0.05 and nano at $0.02 per million tokens. The 1M-token 429 test is unmeasured.
- **Voyage AI embeddings and rerankers (Voyage AI Innovations, Inc.)**: Up to 1,000 texts a call within per-model token caps, a Batch API at a 33 per cent discount, published per-token prices, and output_dtype int8, uint8, binary or ubinary with output_dimension. An account found it the fastest service for 250M tokens.

## 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.

- **Amazon Nova Multimodal Embeddings (Amazon Web Services, Inc.)**: Bedrock batch inference lists the model in us-east-1 and embeddingDimension shortens to 256, 384 or 1024, but a synchronous call embeds one item, there is no int8 or binary output, and the batch price wasn't on a page read here.
- **Cohere Embed and Rerank (Cohere Inc.)**: int8, uint8, binary, ubinary and output_dimension are documented, but calls cap at 96 inputs and 2,000 inputs a minute, Embed Jobs isn't listed for Embed 5, and the client once dropped int8 vectors from later batches.
- **EmbeddingGemma 2 (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Batches encode locally and truncate_dim gives 512, 256 or 128 dimensions with a quality table; an account stored 49.3M items as 256-bit binary codes. No batch route or price exists; the cost and wall-clock time are the owner's hardware.
- **EmbeddingGemma 300M (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Batches encode locally and MRL truncation to 512, 256 or 128 is documented with scores per size, but there is no batch route or price; cost and wall-clock time are the owner's hardware.
- **Gemini Embedding (Google LLC)**: batchEmbedContents and a Batch API at half price ($0.20 standard per million text tokens) with output_dimensionality down to 128, but float output only, standard rate limits only in AI Studio, and the SDK merged N inputs into one vector.
- **Granite Embedding Small English R2 (International Business Machines Corporation)**: Fast to run locally (about 199 documents a second on one H100 per the card), but there is no batch route or price, and no documented way to shorten vectors or return int8 or binary.
- **jina-embeddings-v5-omni-small (Jina AI GmbH, a subsidiary of Elastic N.V. (per the jina-embeddings listing's 2 October 2026 check))**: Lists encode in one call and truncate_dim gives Matryoshka sizes from 32 to 1,024, but there is no batch route or price beyond the hardware, and the non-commercial licence limits who may backfill with it.
- **Mistral Embed and Codestral Embed (Mistral AI (RCS Paris 952 418 325))**: The Batch API accepts /v1/embeddings with custom_id, and codestral-embed takes output_dimension and output_dtype int8, uint8, binary or ubinary; mistral-embed has no output options, and rate limits are only in the admin panel.
- **Nomic Embed (Nomic, Inc.)**: dimensionality shortens nomic-embed-text-v1.5 to 64 to 768, but no public page prices the endpoint, there is no batch route or int8 or binary output, and the Python client sends at most 10 texts a request.
- **NVIDIA NeMo Retriever Embedding and Reranking NIMs (NVIDIA Corporation)**: Lists of inputs per call with dynamic batching, int8 and uint8 output, and free reduced dimensions from 128 to 2,048 on Nemotron 3.5, but no batch route or per-token price; cost is GPUs plus the per-GPU licence.
- **OpenAI embeddings (OpenAI OpCo, LLC)**: Up to 2,048 inputs a request returned with index, and dimensions shortens text-embedding-3 vectors, but output is float or base64 only, and an account found text containing special-token strings failed text-embedding-3-small with HTTP 500.
- **Qwen3-Embedding-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Batches encode locally, but custom dimensions from 32 to 1024 are claimed without saying how to request them, there is no int8 or binary output, and no batch route or price beyond the hardware.
- **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))**: Batches encode locally, but custom dimensions from 32 to 2560 are claimed without saying how to request them, there is no int8 or binary output, and no batch route or price beyond the hardware.
- **Qwen3-Embedding-8B (Alibaba Cloud (named in the licence's copyright line))**: Batches encode locally, but custom dimensions up to 4,096 are claimed without instructions, an account found LM Studio ignored the dimensions parameter and returned 4,096 values, and there is no int8 or binary output.
- **Qwen3-VL-Embedding-2B (Alibaba Cloud (RDAP names Alibaba Cloud (Singapore) Private Limited as the registrant of qwen.ai))**: Batches encode locally and custom dimensions from 64 to 2048 and output quantisation are marked supported, but the card doesn't say how to request either, and there is no batch route or price beyond the hardware.
- **Qwen3-VL-Embedding-8B (Alibaba Group (Tongyi Lab, named in the copyright line of the GitHub repository's Apache 2.0 licence))**: Batches encode locally and custom dimensions from 64 to 4096 and output quantisation are marked supported, but the card doesn't say how to request either, and there is no batch route or price beyond the hardware.
- **ZeroEntropy zerank and zembed (ZeroEntropy, Inc.)**: No hosted batch route remains. The zembed-1 card claims dimensions from 40 to 2,560 and quantisation to binary but doesn't show the parameter, and self-hosting has no published price.

## What stops agents finishing

- SDK merges batches wrongly (Cohere dropped int8 vectors; Gemini returned one vector for N inputs)
- dimensions parameter ignored by the serving route, so every vector is full size
- small per-call caps (Cohere 96 inputs, Nomic client 10 texts) with rate limits only visible in a dashboard
- special-token strings in the text fail the whole request

## Around it

Before: document-extraction.pdf-to-text, data.export-table. After: vector-search.ingest-sync, embeddings.index-docs-for-rag.

## Where this comes from

Weight 12 of 100 in Embeddings & rerankers: 11 of 62 accounts (91 account-use-case tags, accounts counted in each use case they fit, scaled to 100 and rounded by hand); GitHub issues on four SDKs are about batching, plus HN accounts of 49M-item and 250M-token jobs

As data: https://www.anchorterminal.com/api/v1/use-cases/embeddings.bulk-backfill.json, or the MCP tool get_use_case with id embeddings.bulk-backfill. 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
