Support multiple inference engines and model providers via a strategy
enum so each receives its own thinking-disabling field in
chat-completion request bodies.
Strategies:
- "vllm" → chat_template_kwargs: { enable_thinking: false } (vLLM / SGLang)
- "deepseek" → enable_thinking: false (DeepSeek official API)
- "dashscope" → enable_thinking: false (Alibaba DashScope / Qwen)
- "openai" → reasoning_effort: "low" (OpenAI o-series, cannot fully disable)
- "anthropic" → thinking: { type: "disabled" } (Anthropic Claude)
- "kimi" → thinking: { type: "disabled" } (Kimi / Moonshot)
- "gemini" → thinking_config: { thinking_budget: 0 } (Google Gemini)
Defaults to false (no wrapper). Also accepts boolean true as a shorthand
for "vllm" for convenience.
Covered paths:
- StandaloneLLMRunner (OpenClaw plugin + gateway): llm.disableThinking,
wired through parseConfig, tdai-core, seed-runtime, and gateway config
(TDAI_LLM_DISABLE_THINKING env also accepts strategy names).
- Offload local mode (L1/L1.5/L2): separate offload.disableThinking.
The fetch wrapper lives in src/utils/no-think-fetch.ts with a
STRATEGY_TRANSFORMERS map for clean dispatch. StandaloneLLMRunner builds
it once in the constructor; LocalLlmClient caches it at construction
time and passes it through to callLlm().
Add vitest unit tests for all 7 strategies, normalization, validation,
embedding skip, and non-JSON tolerance (20 tests total).
Signed-off-by: jackson.jia <jiazhenghua0@gmail.com>
MAX_RETRIES was hardcoded to 0, making the existing exponential
backoff retry logic dead code. Transient errors (network jitter,
429 rate limits, DNS failures) caused immediate embedding failure,
contributing to JSONL/SQLite drift (issue #156).
Changed MAX_RETRIES from 0 to 3, matching the retry strategy
already used by tcvdb-client.ts.
Closes#159
Signed-off-by: WSL_zhangxiaoshuai <zhangxiaoshuai@wsl.com>
Co-authored-by: WSL_zhangxiaoshuai <zhangxiaoshuai@wsl.com>
* feat(embedding): support ZeroEntropy native embed API
ZeroEntropy's embedding service is reachable but not OpenAI-shaped:
its endpoint is POST /v1/models/embed (not /v1/embeddings), the
request body requires an `input_type` ("query" or "document") and
rejects an explicit `dimensions` field, and the response is wrapped
as `{results: [{embedding}]}` instead of OpenAI's
`{data: [{index, embedding}]}`.
Branch on `provider === "zeroentropy"` inside OpenAIEmbeddingService
rather than introducing a new service class:
- `/models/embed` is used instead of `/embeddings` (also honoured
by the qclaw proxy path so users can still front it).
- request body sets `input_type: "query"` and omits `dimensions`
(zembed-1 is fixed-dimension and rejects explicit overrides).
- response parsing decodes the `results[]` envelope and preserves
input order via array index rather than OpenAI's `index` field.
`openclaw.plugin.json` updates the `embedding.provider` description to
list `zeroentropy` as a recognized value.
Everything else (timeout, retries, batching, char-cap truncation,
sanitize+normalize) is shared with the OpenAI path — only the wire
format changes.
Fixes#68
Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>
* refactor(embedding): split ZeroEntropy into its own EmbeddingService
Addresses @Xuruida's review feedback on PR #137: instead of branching
OpenAIEmbeddingService._callApi on `providerName === "zeroentropy"`,
extract the shared HTTP layer and give ZeroEntropy its own class.
This keeps OpenAIEmbeddingService focused on the OpenAI-compatible
wire format and stops provider-specific protocol differences from
accumulating inside a single class.
Changes:
- New module-level helpers (provider-agnostic):
* `truncateEmbeddingInputs(texts, maxInputChars, logger)` — was
previously a private method duplicated logic-side.
* `postEmbeddingRequest({fetchUrl, headers, body, timeoutMs})` —
owns fetch + AbortController timeout + exponential-backoff
retries + the EmbeddingApiError non-retry rule for 4xx (except
429). Returns the parsed JSON; callers own response parsing.
- `OpenAIEmbeddingService._callApi` is back to OpenAI-only: builds
`{input, model, dimensions?}`, posts to `/embeddings`, parses
`{data: [{index, embedding}]}` with index-sort. The qclaw proxy
branch is unchanged. No more `isZeroEntropy` reads.
- New `ZeroEntropyEmbeddingService implements EmbeddingService`:
- Same `OpenAIEmbeddingConfig` shape (baseUrl / apiKey / model /
dimensions / maxInputChars / timeoutMs are identical on the
wire), but encodes ZeroEntropy's wire format independently:
posts to `${baseUrl}/models/embed`, sends `input_type: "query"`,
omits `dimensions` (zembed-1 is fixed-dim), parses
`{results: [{embedding}]}` preserving array order.
- Doesn't carry the OpenAI-only knobs (`sendDimensions`,
`proxyUrl`, qclaw) so the class surface stays small.
- `createEmbeddingService` factory: new `provider === "zeroentropy"`
branch placed before the OpenAI-compatible catch-all so it wins
the dispatch.
Behaviour preserved:
- All existing OpenAI / DeepSeek / Azure / qclaw deployments hit
exactly the same wire calls as before this PR was opened.
- ZeroEntropy wire format (URL + body + response) matches the
original PR's branched implementation byte for byte.
- Timeout, retry, error mapping, batch splitting, char-cap
truncation, sanitize+normalize are shared via the module helpers
so the two services can't diverge by accident.
Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>
* fix(embedding): forward dimensions to ZeroEntropy when sendDimensions
Addresses @Xuruida's follow-up review on PR #137. The previous commit
silently dropped `dimensions` from every ZeroEntropy request based on
the wrong claim that zembed-1 is fixed-dim.
Verified against
https://docs.zeroentropy.dev/api-reference/models/embed: `dimensions`
is an optional integer; for zembed-1 the accepted set is
[2560, 1280, 640, 320, 160, 80, 40] (Matryoshka truncation). Not
forwarding it left users no way to pick a smaller dim end-to-end —
the wire payload always used the model default while the Float32Array
return type and getDimensions() still reported config.dimensions, so
the runtime vector size and the configured one could disagree.
- ZeroEntropyEmbeddingService now mirrors OpenAIEmbeddingService:
reads `config.sendDimensions ?? true` into a `sendDimensions` field
and emits `dimensions: this.dims` in the request body when set.
- Class docstring corrected — the "rejects explicit overrides" line
is replaced with the documented Matryoshka set. Server still
rejects out-of-set values, but that's a config choice the user
owns; we forward verbatim instead of clamping silently.
Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>
---------
Signed-off-by: 李冠辰 <liguanchen@xiaomi.com>
- Add `recall.maxCharsPerMemory` and `recall.maxTotalRecallChars` with defaults of `0`, which do not alter existing behavior. Users can opt in by setting positive values to cap injected memory context.
- Apply the budget after L1 search and before `<relevant-memories>` injection, preserving score order while truncating oversized entries and dropping overflow.
- Document the new guards in README, README_CN, and `openclaw.plugin.json`.
jackson-jia-914