## What problem does this PR solve?
Go test files are never compiled in CI — only production binaries via
`go build`. This allowed a missing `"sort"` import in
`metadata_filter_test.go` to be merged without detection.
## Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
## Changes
- Add `go test -count=1 ./internal/...` step after Go build in CI
workflow
- Fix missing `"sort"` import in `metadata_filter_test.go` (pre-existing
compile error)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Summary
- Harden `NewNovitaModel` to avoid panics when `http.DefaultTransport`
is a custom non-`*http.Transport` RoundTripper.
- Fallback to a safe transport (`ProxyFromEnvironment`) while preserving
existing pooling/timeout settings.
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
Fixes#15012
The Novita Go driver landed in #14850 and shipped a stub `Rerank` method
that returned `"novita, no such method"`, so Novita could not be used as
a rerank provider in RAGFlow. This PR fills that gap, in the same way
#14895 filled the Embed gap on the same driver.
Novita exposes a public rerank endpoint at `POST
https://api.novita.ai/openai/v1/rerank` that accepts the
Cohere-compatible request shape (`{model, query, documents, top_n}`)
with `Authorization: Bearer <api_key>`. `baai/bge-reranker-v2-m3` is
documented in Novita's model library with a 1024-token limit.
### What problem does this PR solve?
Add a Go driver for Voyage AI (https://voyageai.com), one of the
unchecked providers on the umbrella tracking issue #14736. Voyage AI is
**embed + rerank only** — no chat, no streaming, no `/v1/models`
endpoint. It's the first provider in the Go layer of this shape.
Until this PR, a tenant who configured `voyage` as a model provider in
the Go layer fell through to the default branch of
`internal/entity/models/factory.go` and got the dummy driver.
### What this PR includes
- New `internal/entity/models/voyage.go` with a `VoyageModel`
implementing the `ModelDriver` interface.
- New `conf/models/voyage.json` with 6 embedding models (`voyage-3.5`,
`voyage-3.5-lite`, `voyage-3-large`, `voyage-code-3`, `voyage-law-2`,
`voyage-finance-2`) and 2 rerank models (`rerank-2`, `rerank-2-lite`).
- `factory.go`: route `"voyage"` to `NewVoyageModel`.
- `internal/entity/models/voyage_test.go`: 19 unit tests.
### How the driver works
- **Embed**: `POST /v1/embeddings`. Response is OpenAI-shaped (`{data:
[{embedding, index, object, text}], model, usage}`). Driver reorders by
`index`, rejects duplicate / out-of-range / missing slots, and
short-circuits empty input without an HTTP call.
- **Rerank**: `POST /v1/rerank`. Voyage uses **`top_k`** as the request
param name (not `top_n` like Aliyun/SiliconFlow); the driver translates
`RerankConfig.TopN` → `top_k`. Response is Cohere-shaped (`{data:
[{relevance_score, index}], model}`), so the existing
`RerankResponse{Data: []RerankResult{Index, RelevanceScore}}` shape fits
cleanly.
- **`ListModels`**: returns a hardcoded list of `voyageKnownModels`.
Voyage does **not** expose `/v1/models` (probed live, returns 404), so
the driver synthesizes the list from the same set the config ships. New
upstream models are added by extending one slice.
- **`CheckConnection`**: pings a 1-input embed call against
`voyage-3.5`. Without `/v1/models`, this is the cheapest way to verify
the API key + network path before a tenant tries a real workload.
- **`ChatWithMessages` / `ChatStreamlyWithSender` / `Balance` /
`TranscribeAudio` / `AudioSpeech` / `OCRFile`**: all return `"no such
method"`. Voyage does not host any of these surfaces.
No interface change. No new dependencies.
### How was this tested?
**19 unit tests** in `internal/entity/models/voyage_test.go` — all pass
on go 1.25:
```
$ go test -vet=off -run TestVoyage -count=1 ./internal/entity/models/...
ok ragflow/internal/entity/models 0.036s
```
Coverage: Name; Embed (happy path, reorder, empty-input, missing
key/model, duplicate index, out-of-range index, missing slot); Rerank
(happy path with `top_k` assertion, default-to-len-documents, empty
documents, out-of-range index); ListModels (static list, missing key);
CheckConnection (happy, 401); chat methods sentinels; Balance sentinel;
audio/OCR sentinels.
`go build ./internal/entity/models/...` exits 0.
**Live integration test** against `api.voyageai.com`:
```
=== RUN TestVoyageLiveSmoke
[OK] Name() = "voyage"
[OK] ListModels (static): 8 models -> [voyage-3.5 voyage-3.5-lite voyage-3-large voyage-code-3 voyage-law-2 voyage-finance-2 rerank-2 rerank-2-lite]
[OK] CheckConnection
[OK] Embed vectors=3 dim=1024 indices=[0 1 2]
[OK] Embed(empty) -> 0 vectors
[OK] Rerank results=3 scores=[0.8125 0.59765625 0.39453125]
[OK] ChatWithMessages returns voyage, no such method
[OK] Balance returns voyage, no such method
VOYAGE LIVE SMOKE PASSED
--- PASS: TestVoyageLiveSmoke (0.81s)
```
What the live run proves on the wire:
- Auth (`Bearer <key>`) accepted by `api.voyageai.com`.
- Embed `voyage-3.5` on 3 inputs returns 3 vectors at dim 1024 with
`index` field preserved as `[0, 1, 2]` — the reorder-by-index code is
exercised on real data.
- Empty input short-circuits without an HTTP call (mock server would
have been hit if it did).
- Rerank `rerank-2` on 3 docs returns 3 real `relevance_score` floats
`[0.8125, 0.598, 0.395]`. The `top_k` translation works on the live
wire.
- All sentinel methods return the documented `"no such method"` strings.
### Note on PR history
This branch was previously named for LocalAI Embed work which is now
consolidated into PR #14813. The branch was reset to `upstream/main` and
rebuilt for Voyage. Diff against `main` is a clean +838 lines across 4
files.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
Tracking: #14736
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
Add a Go driver for Novita.ai (https://novita.ai), one of the unchecked
providers on the umbrella tracking issue #14736. Novita exposes an
OpenAI-compatible REST API at `https://api.novita.ai/v3/openai` and
proxies a large catalog of third-party models (DeepSeek, Llama, Qwen3,
Kimi, Gemma, Mistral, MiniMax, GLM, etc.) behind a single OpenAI-shaped
surface — 102 models live at the time of writing.
Until this PR, a tenant who configured `novita` as a model provider in
the Go layer fell through to the default branch of
`internal/entity/models/factory.go` and got the dummy driver.
### What this PR includes
- New `internal/entity/models/novita.go` with a `NovitaModel`
implementing the `ModelDriver` interface (~520 lines).
- New `conf/models/novita.json` with 7 representative chat models
(DeepSeek-V4, Llama-3.3-70B, Qwen3-30B/235B reasoning, Kimi-K2,
Gemma-3-27B, Mistral-Nemo).
- `factory.go`: route `"novita"` to `NewNovitaModel`.
- `internal/entity/models/novita_test.go`: 23 unit tests.
### Notable design point: `<think>...</think>` reasoning extraction
Novita-routed reasoning models like `qwen3-*` and `deepseek-r1-*` embed
their chain-of-thought **inline inside content as `<think>...</think>`
tags**, rather than in a separate `reasoning_content` field. Verified
live by probing `api.novita.ai`:
```
content head 200: <think>
Okay, let's see. I need to find 15% of 80. Hmm, percentages can sometimes be tricky, but I think
content tail 100: h, that works.
Alternatively, 0.15 × 80. If I move the decimal two places to the left for </think>
```
Without handling, a tenant picking qwen3 via Novita would see raw
`<think>` tags in their UI answer — different from every other reasoning
provider in the Go layer.
The driver detects those tags and routes the inner text to
`ChatResponse.ReasonContent` (non-stream) or the sender's second arg
(stream), keeping the visible answer clean of tag clutter:
- **`splitNovitaThink`** — scans a complete content string. Used by the
non-streaming path. Handles multiple `<think>` blocks, unclosed tags
(the model got cut off mid-reasoning), pure-text content with no tags.
- **`novitaThinkExtractor`** — stateful streaming version. Buffers
trailing bytes that might be the start of a tag (e.g. `<thi` held back
when the next chunk completes `nk>`), then emits segments in routing
order so callers can pipe them to a UI. Tested with byte-level chunk
boundaries and tag-spanning scenarios.
### Method coverage
| Method | Behavior |
|---|---|
| `ChatWithMessages` | `POST /v3/openai/chat/completions`, `<think>`
extraction on response |
| `ChatStreamlyWithSender` | SSE stream, stateful `<think>` extraction
across deltas |
| `ListModels` / `CheckConnection` | `GET /v3/openai/models` (102 live)
|
| `Embed` / `Rerank` / `Balance` / `TranscribeAudio` / `AudioSpeech` /
`OCRFile` | `"no such method"` — Novita's OpenAI-compatible surface does
not expose any |
No interface change. No new dependencies.
### How was this tested?
**23 unit tests** in `internal/entity/models/novita_test.go` — all pass:
```
$ go test -vet=off -run "TestNovita|TestSplitNovita" -count=1 ./internal/entity/models/...
ok ragflow/internal/entity/models 0.020s
```
Coverage:
- `splitNovitaThink` (5 cases: pure text, single block, leading text,
multiple blocks, unclosed tag)
- `novitaThinkExtractor` (6 cases: single-chunk, opening tag span,
closing tag span, byte-level chunking, no tags, lone `<` not as tag
start)
- `ChatWithMessages`: pure text, with `<think>` tags, missing API key,
empty messages, HTTP error
- `ChatStreamlyWithSender`: tag-stripping with spanning deltas, pure
content, sender-required, stream-true-required
- `ListModels` / `CheckConnection` (happy paths)
- All sentinel methods
`go build ./internal/entity/models/...` exits 0 on go 1.25.
**Live integration test** against `api.novita.ai/v3/openai`:
```
=== RUN TestNovitaLiveSmoke
[OK] Name() = "novita"
[OK] CheckConnection
[OK] ListModels: 102 models (showing first 6) [deepseek/deepseek-v4-pro deepseek/deepseek-v4-flash deepseek/deepseek-v3.2 xiaomimimo/mimo-v2.5-pro moonshotai/kimi-k2.6 zai-org/glm-5.1]
[OK] Chat (llama-3.3) answer="ok" reason=""
[OK] Chat (qwen3) answer len=0 head=""
ReasonContent len=1657 head="Okay, so I need to figure out what 15% of 80 is. Hmm, percentages can sometimes trip me up, but let ..."
[OK] Stream content: 0 chunks, 0 chars; reasoning: 600 chunks, 1667 chars
[OK] Embed/Rerank/Balance/TranscribeAudio/AudioSpeech/OCRFile all return "novita, no such method"
NOVITA LIVE SMOKE PASSED
--- PASS: TestNovitaLiveSmoke (26.18s)
```
What the live run proves on the wire:
- Auth (`Bearer <key>`) accepted by `api.novita.ai`.
- `/v3/openai/models` parser handles the real 102-model response.
- Non-stream chat against `meta-llama/llama-3.3-70b-instruct`: clean
string answer, empty ReasonContent (non-reasoning model, pure-text
path).
- Non-stream chat against `qwen/qwen3-30b-a3b-fp8`: 1657-char reasoning
extracted from `<think>...</think>` and routed to
`ChatResponse.ReasonContent`. Visible answer is 0 chars in this run
because qwen3 spent its 600-token budget entirely on reasoning before
reaching the answer phase — that's the model's behavior, not a driver
bug. The important thing: **no `<think>` tags leaked into the visible
Answer field**.
- Streaming against qwen3: 600 reasoning chunks (1667 chars) emitted via
the sender's 2nd arg across SSE deltas; **no `<think>` tag fragments
leaked into the content channel** despite tag boundaries crossing chunk
boundaries on the wire.
- All 6 sentinel methods return the documented `"no such method"`
strings.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
Tracking: #14736
