🐛 Bug Report: nomic-embed-text silently truncates input at 512 tokens, and the reported token count is untruncated
还没有人认领这个 Issue。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 45/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- docker, rust
调研方向
Start with src/embedding.rs, where init_model sets the model options, and the tokenizer loading and compute_sub_batch paths described in the report. Check tests/embed_e2e.rs for coverage of vector differences and token counts. The report identifies model context lengths but leaves the default cap and memory tradeoff open; done means embedding respects the chosen cap, reports the count actually embedded, and tests cover both behaviors.
由索引模型根据 Issue 内容生成。
描述
👟 Reproduction steps
- Run
appwrite/embedding:0.3.1(the image pinned in appwrite/appwrite'sdocker-compose.yml) withEMBEDDING_MODELS=nomic-embed-text,all-minilm. - Embed two texts that share a prefix longer than 512 tokens and differ only in the tail, then compare the vectors.
- Embed one long text (
"word " * 20000) and readtokensin the response.
Script
import json, math, urllib.request
def embed(model, texts):
req = urllib.request.Request("http://localhost:3000/embed",
json.dumps({"model": model, "texts": texts}).encode(),
{"content-type": "application/json"})
return json.load(urllib.request.urlopen(req))
def cos(a, b):
return sum(x*y for x, y in zip(a, b)) / math.sqrt(sum(x*x for x in a) * sum(y*y for y in b))
prefix = "fox " * 600 # > 512 tokens
a = embed("nomic-embed-text", [prefix + "cats cats cats cats"])
b = embed("nomic-embed-text", [prefix + "quantum quantum quantum quantum"])
print(cos(a["embeddings"][0], b["embeddings"][0])) # 1.0, the tails were never seen
print(embed("nomic-embed-text", ["word " * 20000])["tokens"]) # 20002
👍 Expected behavior
nomic-embed-text(the default model) embeds up to its real context length (Nomic Embed Text v1.5 supports 8192), or up to a documented, configurable cap.tokensreports the number of tokens the model actually embedded.
👎 Actual Behavior
The model only sees the first 512 tokens and silently ignores the rest. Binary search over the prefix length finds where the tail stops affecting the vector:
| Model | Tail is ignored once the prefix reaches | Cosine of the two vectors |
|---|---|---|
| nomic-embed-text | 512 tokens | 1.000000 |
| all-minilm | 256 tokens (its real limit) | 1.000000 |
At the same time tokens counts the full, untruncated input:
| Model | tokens reported |
Tokens actually embedded |
|---|---|---|
| nomic-embed-text | 20002 | 512 |
| all-minilm | 256 | 256 |
Impact:
- Semantic search over anything longer than ~400 English words matches only on the opening. Two long documents with the same intro get identical vectors, with no error or warning.
- Appwrite's embedding token usage overstates real work by up to ~40x on long inputs.
Root cause:
- Truncation at 512.
init_modelinsrc/embedding.rsbuildsInitOptions::new(model)withoutwith_max_length, so fastembed falls back toDEFAULT_MAX_LENGTH = 512(fastembed-rs/src/text_embedding/mod.rs), and its tokenizer loader applieswith_truncation(max_length). - Untruncated count. Tokens are counted with a separate tokenizer from
load_tokenizer, which readstokenizer.jsondirectly withTokenizer::from_fileand never applies that truncation. nomic'stokenizer.jsoncarries none, so every input token is counted. (minilm's does, which is why its count is right.)
compute_sub_batch also assumes "~1000-2000 token code chunks", which suggests more than 512 tokens was the intent.
💡 Proposed fix
I'd like to take this:
- Explicit context length per model. Pass
.with_max_length(n)ininit_model, wherenis the model's real context (nomic 8192, gemma 2048, bge-small 512, minilm 256), capped by a newEMBEDDING_MAX_TOKENSenv var so self-hosters can trade context for memory. - Count what was embedded. Apply the same
TruncationParams { max_length: n }to the counting tokenizer inload_tokenizer, sotokensis what the model saw. Optionally add atruncatedcount to the response so Appwrite can tell users when input was cut. - Size memory by the real length. Scale
mb_per_textincompute_sub_batchwith the configuredmax_length(attention grows with seq²) instead of the fixed 100/40 MB, so a higher cap can't OOM the container. - Tests. In
tests/embed_e2e.rs: two texts differing only after 600 tokens must produce different nomic vectors;tokensmust never exceedtexts.len() * max_length. - Docs. State each model's limit in
create-text-embeddings.mdin appwrite/appwrite.
Open question: what should the default cap be? 8192 gives nomic its full context but costs much more memory per text. A middle value (for example 2048, in line with the compute_sub_batch comment) covers most real documents. Steps 2 to 4 hold either way.
Related Appwrite endpoint issues: https://github.com/appwrite/appwrite/issues/15532
🎲 Appwrite version
Different version (specify in environment)
💻 Operating system
MacOS
🧱 Your Environment
appwrite/embedding:0.3.1 in Docker on macOS (CPU, ONNX), as pinned by appwrite/appwrite main at c7949c2887 (12 Nov 2026). The relevant code is unchanged on this repo's main (0.3.3): no commit since 0.3.1 touches max_length or the counting tokenizer.
👀 Have you spent some time to check if this issue has been raised before?
- I checked and didn't find similar issue
🏢 Have you read the Code of Conduct?
- I have read the Code of Conduct
- 主要语言
- Rust
- 星标
- 0
- 派生
- 3
- 平均合并
- 10 小时 33 分钟
- 30 天内合并 PR
- 4
环境准备
- 提供 Dockerfile 或 Docker Compose 文件
- 没有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
appwrite/embedding 的其他 Issue
查看 appwrite/embedding 的全部 Issue
相似的 Issue
-
marketing
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 88/100
flashbots/panetiere#46 · 1 条评论 ·
维护者通常 5 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
jlcodes99/cockpit-tools#3190 ·
维护者通常 4 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
Feyisara2108/stellargive#1376 ·
维护者通常 1 天内回复
-
area:post bug effort:medium P7 source:agent
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复