Is this feature related to a problem? | 该功能需求是否与某个问题相关？请描述
Support ZeroEntropy native embedding API in memory-tencentdb
Summary
Please add first-class support for ZeroEntropy as an embedding provider in @tencentdb-agent-memory/memory-tencentdb.
Currently, memory-tencentdb treats non-local embedding providers as OpenAI-compatible and calls:
ts
${this.baseUrl}/embeddings
This works for OpenAI-compatible providers, but ZeroEntropy’s working embedding endpoint is native/non-OpenAI-shaped:
POST https://api.zeroentropy.dev/v1/models/embed
with model:
zembed-1
and request body shape:
json
{
"input": ["text to embed"],
"input_type": "query",
"model": "zembed-1"
}
The response shape is also different from OpenAI:
json
{
"results": [
{
"embedding": [...]
}
]
}
rather than:
json
{
"data": [
{
"index": 0,
"embedding": [...]
}
]
}
Environment
- Package:
@tencentdb-agent-memory/memory-tencentdb
- Observed version:
0.3.5
- Runtime: Hermes sidecar gateway
- Gateway health reports:
vectorStore: trueembeddingService: true
- Embedding config:
provider: openaibaseUrl: https://api.zeroentropy.devmodel: zembed-1dimensions: 2560
Current behavior
With ZeroEntropy configured, memory-tencentdb calls:
POST https://api.zeroentropy.dev/embeddings
That returns:
HTTP 404
If /v1 is included in the base URL, the OpenAI-style endpoint still fails:
POST https://api.zeroentropy.dev/v1/embeddings
HTTP 404 {"detail":"Not Found"}
As a result, the embedding service object exists, but no vectors are persisted.
Example local DB state after multiple captured conversations:
l0_conversations: 123
l1_records: 3
l0_vec_rowids: 0
l1_vec_rowids: 0
Memory search falls back to FTS for exact keyword matches, while semantic/vector search returns no results.
Expected behavior
memory-tencentdb should support ZeroEntropy embeddings, either via:
- A dedicated provider, e.g.
yaml
embedding:
enabled: true
provider: zeroentropy
baseUrl: https://api.zeroentropy.dev
model: zembed-1
dimensions: 2560
or
- A configurable embedding endpoint/response adapter that can support non-OpenAI-compatible APIs.
For ZeroEntropy, the provider should:
POST {baseUrl}/v1/models/embed
json
{
"input": ["..."],
"input_type": "query",
"model": "zembed-1"
}
results[].embedding
- return normalized
Float32Array embeddings as the existing embedding service does.
Validation
Direct ZeroEntropy native call succeeds:
POST https://api.zeroentropy.dev/v1/models/embed
model: zembed-1
Result:
HTTP 200
embedding dimensions: 2560
But the current OpenAI-compatible path used by memory-tencentdb fails:
POST https://api.zeroentropy.dev/embeddings
HTTP 404
Describe the solution you'd like | 描述你期望的解决方案
Suggested implementation
Add a ZeroEntropyEmbeddingService or branch inside the existing remote embedding service:
ts
if (providerName === "zeroentropy") {
const fetchUrl = ${baseUrl.replace(//+$/, "")}/v1/models/embed;
const body = {
input: texts,
input_type: "query",
model: this.model,
};
const json = await resp.json();
if (!json.results || !Array.isArray(json.results)) {
throw new Error("ZeroEntropy embedding API returned unexpected format: missing 'results' array");
}
return json.results.map((r) => sanitizeAndNormalize(r.embedding));
}
Optionally allow input_type to be configured (query vs document) if users need both retrieval modes.
Why this matters
ZeroEntropy provides 2560-dimensional embeddings through zembed-1, and works well as a hosted embedding provider. Supporting it directly would let Hermes/TencentDB users use ZeroEntropy for semantic memory search without needing an OpenAI-compatible proxy.
Describe alternatives you've considered | 描述你考虑过的其他方案
No response
Additional context | 补充说明
No response
Is this feature related to a problem? | 该功能需求是否与某个问题相关？请描述
Support ZeroEntropy native embedding API in
memory-tencentdbSummary
Please add first-class support for ZeroEntropy as an embedding provider in
@tencentdb-agent-memory/memory-tencentdb.Currently,
memory-tencentdbtreats non-local embedding providers as OpenAI-compatible and calls:ts
${this.baseUrl}/embeddings
This works for OpenAI-compatible providers, but ZeroEntropy’s working embedding endpoint is native/non-OpenAI-shaped:
POST https://api.zeroentropy.dev/v1/models/embed
with model:
zembed-1
and request body shape:
json
{
"input": ["text to embed"],
"input_type": "query",
"model": "zembed-1"
}
The response shape is also different from OpenAI:
json
{
"results": [
{
"embedding": [...]
}
]
}
rather than:
json
{
"data": [
{
"index": 0,
"embedding": [...]
}
]
}
Environment
@tencentdb-agent-memory/memory-tencentdb0.3.5vectorStore: trueembeddingService: trueprovider: openaibaseUrl: https://api.zeroentropy.devmodel: zembed-1dimensions: 2560Current behavior
With ZeroEntropy configured,
memory-tencentdbcalls:POST https://api.zeroentropy.dev/embeddings
That returns:
HTTP 404
If
/v1is included in the base URL, the OpenAI-style endpoint still fails:POST https://api.zeroentropy.dev/v1/embeddings
HTTP 404 {"detail":"Not Found"}
As a result, the embedding service object exists, but no vectors are persisted.
Example local DB state after multiple captured conversations:
l0_conversations: 123
l1_records: 3
l0_vec_rowids: 0
l1_vec_rowids: 0
Memory search falls back to FTS for exact keyword matches, while semantic/vector search returns no results.
Expected behavior
memory-tencentdbshould support ZeroEntropy embeddings, either via:yaml
embedding:
enabled: true
provider: zeroentropy
baseUrl: https://api.zeroentropy.dev
model: zembed-1
dimensions: 2560
or
For ZeroEntropy, the provider should:
POST {baseUrl}/v1/models/embed
json
{
"input": ["..."],
"input_type": "query",
"model": "zembed-1"
}
results[].embedding
Float32Arrayembeddings as the existing embedding service does.Validation
Direct ZeroEntropy native call succeeds:
POST https://api.zeroentropy.dev/v1/models/embed
model: zembed-1
Result:
HTTP 200
embedding dimensions: 2560
But the current OpenAI-compatible path used by
memory-tencentdbfails:POST https://api.zeroentropy.dev/embeddings
HTTP 404
Describe the solution you'd like | 描述你期望的解决方案
Suggested implementation
Add a
ZeroEntropyEmbeddingServiceor branch inside the existing remote embedding service:ts${baseUrl.replace(//+$ /, "")}/v1/models/embed;
if (providerName === "zeroentropy") {
const fetchUrl =
const body = {
input: texts,
input_type: "query",
model: this.model,
};
const json = await resp.json();
if (!json.results || !Array.isArray(json.results)) {
throw new Error("ZeroEntropy embedding API returned unexpected format: missing 'results' array");
}
return json.results.map((r) => sanitizeAndNormalize(r.embedding));
}
Optionally allow
input_typeto be configured (queryvsdocument) if users need both retrieval modes.Why this matters
ZeroEntropy provides 2560-dimensional embeddings through
zembed-1, and works well as a hosted embedding provider. Supporting it directly would let Hermes/TencentDB users use ZeroEntropy for semantic memory search without needing an OpenAI-compatible proxy.Describe alternatives you've considered | 描述你考虑过的其他方案
No response
Additional context | 补充说明
No response