Declare doc_id, filename, mime_type, and size as separate outputs on the
Document Generation component so downstream nodes (e.g., the Code
component) can consume them via the variable picker. The existing
download JSON blob is preserved unchanged for the Message component's
download-chip rendering.
### What problem does this PR solve?
The Document Generation component previously exposed only a single
`download` output —
a JSON-encoded blob containing the file's `doc_id`, `filename`,
`mime_type`, `size`,
and base64 payload. On top of that, the variable picker actively hides
this `download`
entry from every consumer except the Message component (because the
embedded base64 is
too heavy to splat into arbitrary downstream nodes).
The combined effect: users wiring the Doc Generator's output into a Code
component had
no way to retrieve basic file info such as `file_name` or `doc_id` from
the picker,
blocking workflows that need to post-process the generated file (e.g.,
registering it
elsewhere, custom delivery, follow-up API calls).
This PR declares `doc_id`, `filename`, `mime_type`, and `size` as
**discrete outputs**
on the Document Generation component, alongside the existing `download`
blob. The new
fields:
- Appear in the variable picker for **all** downstream nodes, including
the Code
component, so users can bind them directly to script arguments.
- Are cheap scalars only — no base64 payload leaks into other
components.
- Leave the existing `download` JSON blob completely untouched, so the
Message
component's download-chip rendering (which parses that blob via
`_is_download_info`)
keeps working with no behavior change.
Changes:
- `agent/component/docs_generator.py` — declare the four new outputs in
`DocGeneratorParam` and emit them via `set_output(...)` in `_invoke`.
- `web/src/pages/agent/constant/index.tsx` — extend
`initialDocGeneratorValues.outputs`
with the new keys.
- `web/src/pages/agent/form/doc-generator-form/index.tsx` — mirror the
new outputs in
the zod schema so the form is valid.
No changes needed to the picker's existing `download`-hiding filter — it
matches only
on the literal output name `download`, so the new metadata entries fall
through
naturally.
Reported in: https://github.com/infiniflow/ragflow/issues/14461.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
Implement embed for Tencent Hunyuan
**Verified from CLI**
```
RAGFlow(user)> embed text 'what is rag' 'who are you' with 'hunyuan-embedding@test1@hunyuan' dimension 16;
+-----------+-------+
| dimension | index |
+-----------+-------+
| 1024 | 0 |
| 1024 | 1 |
+-----------+-------+
```
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
- [x] Refactoring
### What problem does this PR solve?
1. Fix /chat/completions to send only the latest message
2. Allo chat stream=False
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Go: implement provider: PaddleOCR_Local
**Verified from CLI**
```
RAGFlow(user)> ocr with 'PaddleOCR-VL@test@paddleocr_local' file './internal/test1.jpg'
+----------------------+
| text |
+----------------------+
| ## Parallel to these |
+----------------------+
```
### Type of change
- [X] Bug Fix (non-breaking change which fixes an issue)
- [X] New Feature (non-breaking change which adds functionality)
- [X] Refactoring
## Summary
- Adds a `Hunyuan` Go driver so the new API server can route Tencent
Hunyuan chat instances (registered in `conf/llm_factories.json:3830` as
`Tencent Hunyuan`). Follows the same SaaS-driver shape used for
Astraflow, Avian, Novita, TogetherAI, Replicate, DeepInfra, Upstage, and
LongCat.
Closes#15087
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
Fix /chat/completions not aware of conversation_id
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Closes#15102.
OpenAI's Go provider config advertises `whisper-1` as ASR and `tts-1` as
TTS, but the Go driver returned `openai, no such method` for both audio
paths and did not define `url_suffix.asr` / `url_suffix.tts`.
This PR:
- adds OpenAI audio URL suffixes for `audio/transcriptions` and
`audio/speech`
- implements non-streaming `TranscribeAudio` using multipart form
uploads
- implements non-streaming `AudioSpeech` using the OpenAI speech JSON
request shape
- keeps streaming TTS explicitly unsupported instead of sending binary
audio through the text SSE sender
- adds focused tests for config coverage, ASR/TTS request shape,
required TTS voice validation, and unsupported streaming TTS
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
Fix: /openai/<chat_id>/chat/completions not aware of session_id
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Go: implement reasoning_chat, TTS, ASR for Groq
**Verify from CLI**
```
RAGFlow(user)> think chat with 'qwen/qwen3-32b@test@groq' message 'who r u'
Thinking: Okay, the user asked, who r u. I need to determine what the user is asking. They may be asking about my identity. I should introduce my name and basic functions. The user might want to know what I can do, so I should list some common use cases, such as answering questions, creating writing, coding, and expressing opinions. The user may be curious about how they can interact with me, so they can be advised to ask any questions or provide instructions. Keep your answers conversational, avoid overly technical terms, keep answers concise, and encourage further interaction. Check if there's any ambiguity in the answer and make sure it's accurate and meets the user's needs. Also consider if there are other aspects the user may be interested in, such as my training data or performance. But since the question is basic, I'll focus on the essentials first and invite the user to ask more. In summary, respond to the user's questions by introducing yourself, your functions, and encouraging further interaction.
Answer: Hello! I'm Qwen. I am a large-scale language model developed by Tongyi Lab, designed to assist you in various ways, such as answering questions, creating text, logical reasoning, programming, and more. I aim to provide clear, accurate, and helpful information and support. How can I assist you today? Feel free to ask any questions or give me tasks! 😊
Time: 2.199908
RAGFlow(user)> stream think chat with 'openai/gpt-oss-20b@test@groq' message 'who r u'
Thinking: to respond politely.
Answer: ’m ChatGPT—an AI language model created by OpenAI. I’m here to answer questions, offer explanations, and help with a wide range of topics. How can I assist you today?
RAGFlow(user)> tts with 'canopylabs/orpheus-arabic-saudi@test@groq' text 'hello? show yourself' play format 'wav' param '{"voice": "fahad"}'
SUCCESS
RAGFlow(user)> asr with 'whisper-large-v3-turbo@test@groq' audio './internal/test.wav' param '{"language": "en"}'
+----------------------------------------------------------------------------------------------------------------------+
| text |
+----------------------------------------------------------------------------------------------------------------------+
| The examination and testimony of the experts enabled the Commission to conclude that five shots may have been fired |
+----------------------------------------------------------------------------------------------------------------------+
```
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
Closes#15088.
Adds Groq support to the Go model-provider layer so Groq instances can
be routed through the Go API server with the same OpenAI-compatible
chat, streaming, model listing, and connection-check flow used by other
SaaS providers.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
## Summary
- Added a Groq Go model driver.
- Added the Groq provider catalog and default OpenAI-compatible API URL.
- Registered Groq in the model factory.
- Added focused provider tests.
## What changed
- Implemented chat completions, SSE streaming, ListModels, and
CheckConnection for Groq.
- Covered request shape, stream termination, reasoning fallback, model
listing, custom base URLs, safe transport setup, and unsupported
methods.
- Kept the provider catalog scoped to current Groq chat-capable model
IDs.
- Cleaned up pre-existing Go model package validation blockers so the
package can be tested normally with vet enabled.
## Why
The existing Python/provider catalog path includes Groq, but the Go
model-provider layer did not have a Groq driver, so the Go API server
could not instantiate or use Groq as requested in #15088.
## Notes
The model package now validates without disabling vet.
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
move agent attachment download api to the correct route and update
frontend callers
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### Notes
- Move the attachment download endpoint from document routes to agent
routes.
- Update frontend download callers to use the agent attachment endpoint.
- Reuse the shared file response header helper instead of duplicating it
in `agent_api.py`.
## Summary
- Adds a `TokenPony` Go driver so the new API server can route TokenPony
chat instances, matching the existing Python `TokenPonyChat`
(`rag/llm/chat_model.py:1210`). Follows the same SaaS-driver shape used
for Astraflow, Avian, Novita, TogetherAI, Replicate, DeepInfra, Upstage,
and LongCat.
Closes#15086
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
## Summary
Implements the TODO in `evaluation_service.py`: **Track token usage** in
evaluation results.
## Changes
- **Import** `num_tokens_from_string` from `common.token_utils`
- **Prompt tokens**: Use the full prompt returned by `async_chat` when
available (includes system prompt + knowledge base + query), otherwise
fall back to the question token count
- **Completion tokens**: Count tokens in the generated answer
- **Storage**: Store `token_usage` as `{prompt_tokens,
completion_tokens, total_tokens}` in each `EvaluationResult` instead of
`None`
## Why
The evaluation pipeline previously saved `token_usage: None` for every
result. This change allows downstream consumers (e.g. evaluation
dashboards, cost tracking) to see approximate token usage per test case
using the same tokenizer (tiktoken cl100k_base) used elsewhere in
RAGFlow.
## Testing
- No new tests added; existing evaluation flow unchanged
- Token counting uses existing `num_tokens_from_string` utility
---------
Co-authored-by: kiannidev <kiannidev@users.noreply.github.com>
### What problem does this PR solve?
Fixes#15066
OpenRouter now exposes an official speech-to-text endpoint at `POST
/api/v1/audio/transcriptions`, but the Go model driver still returned
`openrouter, no such method` from `TranscribeAudio`. This left
OpenRouter ASR models unavailable through the Go API server even though
the provider already has OpenRouter audio support for TTS.
Related provider-tracking context: #14736
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
The agent API currently does not pass chat_template_kwargs to the
underlying LLM call path, so clients cannot control template-level model
behavior (such as thinking-mode toggles) when invoking
/agents/chat/completion. This PR adds passthrough support for
chat_template_kwargs across agent execution flows (session and
non-session, streaming and non-streaming) by propagating it through
canvas runtime state and into LLM invocation kwargs. This addresses the
feature gap raised in [Issue
#14182](https://github.com/infiniflow/ragflow/issues/14182).
Closes#14182
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
Closes#14789
### What problem does this PR solve?
User API endpoints (`login`, `user_profile`, `user_add`,
`forget_reset_password`) were returning full user objects via
`to_json()` / `to_dict()`, which included sensitive fields like
`password` and `access_token` in the response body. This leaks
credentials to the client.
This PR adds a `to_safe_dict()` method on the `User` model that strips
sensitive fields (`password`, `access_token`) and replaces all affected
call sites to use it.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
1. Enhance retry and timeout, and adjust the default timeout
2. NER: spacy do not batch chunks
3. extract _has_cancel_and_exit
4. enhance log messages
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
- [x] Refactoring
Closes#4310
### What problem does this PR solve?
Issue #4310 requests practical examples for the RAGFlow SDK and HTTP API
to help developers get started faster. The existing `example/sdk/`
folder only contains `dataset_example.py`. This PR fills the remaining
gaps by adding examples for three key API areas not yet covered in
`main` or by other open PRs (#13904, #13284):
- **Chunk management** — add, list, update, delete, and retrieve chunks
within a dataset
- **Chat assistant** — create a chat assistant, open a session, send
messages (streaming and non-streaming), and clean up
- **Retrieval** — perform semantic retrieval across one or multiple
datasets
### Type of change
- [x] Documentation Update
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
Closes#14865
`download_img` in `common/misc_utils.py` is used for OAuth avatar URLs.
The previous implementation called `async_request` from
`common.http_client`, which followed redirects without re-validating
each hop and did not apply the same SSRF protections as this path needs.
That made it possible to reach non-public or disallowed targets (for
example via redirects or unsafe URLs) when fetching avatars.
This change replaces that flow with an explicit, bounded fetch: each URL
(including every redirect target) is checked with
`common.ssrf_guard.assert_url_is_safe`, DNS is pinned with
`pin_dns_global`, `httpx` streams the body with `follow_redirects=False`
and a manual redirect loop (capped by
`RAGFLOW_OAUTH_AVATAR_MAX_REDIRECTS`), and total response size is capped
(`RAGFLOW_OAUTH_AVATAR_MAX_BYTES`). Timeouts, proxy, and user agent
align with `HTTP_CLIENT_*` env vars without importing `http_client`, so
lightweight tests stay simple.
Unit tests cover empty/None URLs, loopback, cloud metadata-style
addresses, and disallowed schemes so SSRF regressions are caught early.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
---------
Co-authored-by: Kevin Hu <kevinhu.sh@gmail.com>
### What problem does this PR solve?
This PR implements ASR and TTS support for the ZhipuAI Go driver.
The ZhipuAI model config already advertises `glm-asr-2512` as an ASR
model, but the Go driver returned `zhipu, no such method` from
`TranscribeAudio`. This adds the documented audio transcription endpoint
suffix and sends multipart transcription requests with `model`,
`stream=false`, and `file` fields.
Per maintainer review, this also adds the ZhipuAI TTS endpoint suffix
and implements `AudioSpeech` / `AudioSpeechWithSender` for `glm-tts`.
Closes#15133
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
Closes#15089.
Adds PPIO support to the Go model-provider layer so PPIO instances can
be routed through the Go API server with the same OpenAI-compatible
chat, streaming, model listing, and connection-check flow used by other
SaaS providers.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
## Summary
- Added a PPIO Go model driver.
- Added the PPIO provider catalog and default OpenAI-compatible API URL.
- Registered PPIO in the model factory.
- Added focused provider and provider-manager tests.
## What changed
- Implemented chat completions, SSE streaming, ListModels, and
CheckConnection for PPIO.
- Covered request shape, stream termination, reasoning fallback, model
listing, custom base URLs, safe transport setup, unsupported methods,
and provider config loading.
- Kept the provider catalog aligned with the existing RAGFlow PPIO
factory model set.
- Cleaned up pre-existing Go model package validation blockers so the
scoped provider tests can run normally with vet enabled.
## Why
The existing Python/provider catalog path includes PPIO, but the Go
model-provider layer did not have a PPIO driver, so the Go API server
could not instantiate or use PPIO as requested in #15089.
### What problem does this PR solve?
implement rerank, asr, tts for TogetherAI
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
1. update python version to 3.13
2. upgrade ormsgpack to 1.6.0
### Type of change
- [x] Refactoring
---------
Signed-off-by: Jin Hai <haijin.chn@gmail.com>
### What problem does this PR solve?
implement ASR and TTS for Xinference
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
- [x] Refactoring
## Summary
Fixes 10 unguarded `response.choices[0]` accesses that cause
`IndexError` or `AttributeError` when the LLM returns an empty `choices`
list — the scenario described in #14711.
- `rag/llm/cv_model.py`
- `rag/llm/chat_model.py`
Each access site is now guarded with:
```python
if not response.choices:
raise ValueError("LLM returned empty response")
```
## Verification
Detected and verified by [pact](https://github.com/qizwiz/pact) — a
sheaf-cohomological LLM contract checker using Z3 as a local theory
solver.
**pact sheaf-cohomological proof status after fix:**
| File | Ȟ¹ (after) | Z3 |
|------|-----------|-----|
| `rag/llm/cv_model.py` | 0 | UNSAT ✓ |
| `rag/llm/chat_model.py` | 0 | UNSAT ✓ |
All access sites proven safe (Z3 UNSAT certificate).
The checker was also used to verify the autogen streaming-None fix in
[microsoft/autogen#7711](https://github.com/microsoft/autogen/pull/7711).
## Test plan
- [ ] Existing test suite passes
- [ ] Manually test with a provider that returns empty `choices` under
load (e.g. Vertex AI)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: Jonathan Hill <jonathan.f.hill@gmail.com>
`GET /agents/<agent_id>/sessions/<session_id>` crashed with
`AttributeError: 'NoneType' object has no attribute 'to_dict'` when the
session lookup failed: `_, conv =
API4ConversationService.get_by_id(...)` returned `(False, None)`, then
`conv.to_dict()` was called unconditionally.
This is reachable in multi-instance deployments: the session row may not
yet be visible on the node servicing the immediate follow-up GET after a
session is created on a different node.
Add the same `if not exists` guard already used by every other call site
of `API4ConversationService.get_by_id` (see agent_api.py:1147,
sdk/session.py:179, conversation_service.py:248, canvas_service.py:323).
Closes#14989
### What problem does this PR solve?
_Briefly describe what this PR aims to solve. Include background context
that will help reviewers understand the purpose of the PR._
### Type of change
- [ ] Bug Fix (non-breaking change which fixes an issue)
- [ ] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
Replace the RuntimeError with a warning + first-address fallback so a
single email whose From header contains multiple addresses no longer
crashes the entire IMAP sync task. Also add regression tests covering:
- #14963: RFC 5322 quoted display names with commas (e.g. "Schlüter,
Sabine" <s@x>) parsed as one address, not two.
- #14964: multi-address headers warn instead of raising.
Closes#14964
Refs #14963
## Summary
- Bump pinned nginx in `Dockerfile` from `1.29.5-1~noble` (vulnerable)
to `1.31.0-1~noble` to remediate **CVE-2026-42945**.
## Root Cause
`Dockerfile:58` pinned `ARG NGINX_VERSION=1.29.5-1~noble`. Per the
official nginx security advisory, **CVE-2026-42945** is a buffer
overflow in `ngx_http_rewrite_module` triggered via the `rewrite` and
`set` directives, affecting nginx **0.6.27 through 1.30.0**. `1.29.5`
falls inside that range, so the shipped image is vulnerable.
References:
- nginx security advisories:
https://nginx.org/en/security_advisories.html
- Vendor advisory: https://my.f5.com/manage/s/article/K000161019
- Fixed versions: `1.31.0` (mainline) and `1.30.1` (stable)
## Fix
Single-line change in `Dockerfile:58`:
```diff
-ARG NGINX_VERSION=1.29.5-1~noble
+ARG NGINX_VERSION=1.31.0-1~noble
### What problem does this PR solve?
Fixes#14997.
RAPTOR builds on the Infinity backend have been broken since v0.25.2
introduced the `extra` field in code (`rag/svr/task_executor.py:1011`)
without declaring it in `conf/infinity_mapping.json`. Every RAPTOR job
fails with:
```
infinity.common.InfinityException: (3013, 'Fail to bind the expression: extra@src/planner/expression_binder_impl.cpp:99')
```
The auto-migration in
`common/doc_store/infinity_conn_base.py:_migrate_db()` adds any columns
it finds in the mapping JSON to existing tables — so the only thing
standing between users and a working RAPTOR build is that one missing
declaration. OceanBase, ES, and OpenSearch were unaffected because they
store `extra` as a native JSON type; only Infinity (which has a strict
`varchar`/`integer`/`float` schema) needed the addition.
### The fix
Two-part change:
1. **`conf/infinity_mapping.json`**: declare `"extra": {"type":
"varchar", "default": ""}`. On next startup, `_migrate_db()` adds the
column to all existing chunk tables — no manual DDL needed for upgrading
installations.
2. **`rag/utils/infinity_conn.py` `insert()`**: serialize the `extra`
dict to a JSON string at write time, since Infinity's `varchar` can't
store a Python dict directly. Modelled on the existing `chunk_data`
handling a few lines above.
The read path (`rag/utils/raptor_utils.py:_as_extra_dict`) already
normalises both dict and JSON-string inputs, so no read-side change is
needed. Other backends are untouched — `task_executor.py` still writes
the dict, and the OceanBase/ES/OpenSearch insert paths handle dicts
natively.
### Verification
Tested on a v0.25.4 deployment with the Infinity backend by applying the
same two changes via mounted-volume override:
- Confirmed `_migrate_db()` adds the `extra` column to all pre-existing
chunk tables on startup (column visible via Infinity's
`show_columns()`).
- Triggered RAPTOR builds on four datasets (~21k chunks total) via `POST
/api/v1/datasets/<id>/index?type=raptor`.
- All four progressed past the previously-failing
`get_raptor_chunk_methods()` call into actual entity-extraction and
clustering work without the (3013) error.
- GraphRAG builds (which can trigger the same path indirectly via
`task_executor.py:857`) also progressed cleanly.
### Type of change
- [X] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
`UpstageModel.ChatStreamlyWithSender` (in the driver merged via #14819)
only extracted `delta.content` from each SSE event. For the `solar-pro3`
reasoning family (and any future Upstage model that follows the same
wire shape), the chain-of-thought is streamed in a **separate
`delta.reasoning` field**, and the driver was silently dropping all of
it.
The non-streaming path already extracts `message.reasoning` into
`ChatResponse.ReasonContent` (added earlier in this PR's history), so
the same model produced **inconsistent behavior** between streaming and
non-streaming: a tenant calling `solar-pro3` with `reasoning_effort:
high` would see the reasoning trace if they used `ChatWithMessages` but
not if they used `ChatStreamlyWithSender`.
### Live evidence
Probed against `api.upstage.ai/v1/chat/completions` with `solar-pro3` +
`reasoning_effort: high` + `stream: true` (8000-token budget so the
reasoning has room to finish):
```
$ curl -sN -H "Authorization: Bearer <key>" -H "Content-Type: application/json" \
-X POST https://api.upstage.ai/v1/chat/completions \
-d '{"model":"solar-pro3","messages":[{"role":"user","content":"Compute 15% of 80."}],
"max_tokens":8000,"stream":true,"reasoning_effort":"high"}'
# across 168 SSE events:
# delta keys seen: [content reasoning role]
# delta.content total len: 121 chars (the visible answer)
# delta.reasoning total len: 159 chars (the chain-of-thought) <- driver dropped this
```
A representative event showing both fields side by side:
```json
data: {"choices":[{"index":0,"delta":{"reasoning":"15% = 0.15."}}]}
data: {"choices":[{"index":0,"delta":{"content":"15% of 80 is "}}]}
```
The 159 chars of reasoning were arriving on the wire and being thrown
away. `solar-pro2` was also probed (625 events); it does **not** emit
`delta.reasoning` — its reasoning is inlined into `delta.content` — so
this change is a no-op for it and for `solar-mini`.
### What this PR includes
- `internal/entity/models/upstage.go`: in the SSE scanner loop, extract
`delta.reasoning` before `delta.content` and forward each non-empty
chunk via the sender's second arg (the existing `reasonContent` channel
the non-stream path already populates).
The ordering contract is documented inline: reasoning chunks within a
single SSE event are emitted before content chunks, so a UI that pipes
both sees the chain-of-thought start before the answer for that token,
matching the wire order Upstage emits.
- `internal/entity/models/upstage_test.go`: three new tests pinning the
new behavior:
- `TestUpstageStreamExtractsReasoningDelta` — reasoning + content
forwarded to the right sender args; one-of invariant per call
- `TestUpstageStreamReasoningChunksArriveBeforeContent` — ordering
pinned within a single SSE event that carries both fields
- `TestUpstageStreamWithoutReasoningStillWorks` — regression net:
non-reasoning models (`solar-mini`, `solar-pro2`) continue to work; the
reason callback never fires
No interface change. No factory change. No config change.
### How was this tested?
```
$ go test -vet=off -run TestUpstage -count=1 -v ./internal/entity/models/...
... (existing tests 1..9 still pass) ...
=== RUN TestUpstageStreamExtractsReasoningDelta
--- PASS: TestUpstageStreamExtractsReasoningDelta (0.01s)
=== RUN TestUpstageStreamReasoningChunksArriveBeforeContent
--- PASS: TestUpstageStreamReasoningChunksArriveBeforeContent (0.01s)
=== RUN TestUpstageStreamWithoutReasoningStillWorks
--- PASS: TestUpstageStreamWithoutReasoningStillWorks (0.00s)
PASS
ok ragflow/internal/entity/models 0.034s
```
12/12 Upstage tests pass on go 1.25. `go build
./internal/entity/models/...` exits 0.
**Live integration test** (smoke test not committed) — the patched
driver was run directly against `api.upstage.ai/v1` with the same prompt
that produced the curl evidence above:
```
=== RUN TestUpstageStreamReasoningLiveSmoke
[OK] visible content: 50 chunks, 84 chars
[OK] reasoning: 39 chunks, 90 chars
content head 200: "\\(15\\% = \\frac{15}{100}=0.15\\).\n\n\\[\n0.15 \\times 80 = 12.\n\\]\n\n**15 % of 80 is 12.**"
reasoning head 200: "We need to compute 15% of 80. That's 0.15 * 80 = 12. So answer is 12. Provide explanation."
UPSTAGE STREAM REASONING SMOKE PASSED
--- PASS: TestUpstageStreamReasoningLiveSmoke (1.97s)
```
Before this fix, the same call would have produced **0 reasoning
chunks**. The 90 chars of reasoning that the patched driver now surfaces
are the chain-of-thought solar-pro3 emits when reasoning_effort is high.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
`MistralModel.ChatWithMessages` (in the driver merged via #14807)
assumes that `choices[0].message.content` from `/v1/chat/completions` is
always a string and falls through to `return nil, fmt.Errorf("invalid
content format")` on anything else.
That assumption breaks for the **magistral reasoning family**
(`magistral-small-*`, `magistral-medium-*`). When the model needs a
chain-of-thought to answer, Mistral returns `content` as a **structured
array of typed parts**:
```json
"content": [
{"type": "thinking",
"thinking": [{"type": "text", "text": "Combined speed is 150 mph. 300 / 150 = 2 hours."}],
"closed": true},
{"type": "text", "text": "They will meet after **2 hours**."}
]
```
Concretely, this is what the live API returns today (probed against
`api.mistral.ai/v1`):
```
$ curl -H "Authorization: Bearer <key>" -H "Content-Type: application/json" \
-X POST https://api.mistral.ai/v1/chat/completions \
-d '{"model":"magistral-medium-latest",
"messages":[{"role":"user","content":"two trains 60mph and 90mph, 300mi apart, when do they meet? step by step."}],
"max_tokens":1024}'
HTTP 200
{ "choices":[{"message":{
"role":"assistant",
"content":[
{"type":"thinking","thinking":[{"type":"text","text":"Okay, let's see..."}],"closed":true},
{"type":"text","text":"To determine when the two trains meet..."}
]}}] }
```
With the current driver, every call like that returns the generic
`"invalid content format"` error. Trivial prompts that happen to fit in
a string answer still succeed, so the breakage is **non-deterministic
from the tenant's POV**: same model, same provider, sometimes works,
sometimes 500s with no useful error.
A secondary issue: `conf/models/mistral.json` does not include any
magistral model. The picker hid the broken path, which is why this
wasn't caught during #14807's review.
### What this PR includes
- New helper `extractMistralContent(raw interface{}) (answer,
reasonContent string, err error)` in
`internal/entity/models/mistral.go`, which normalizes both shapes
Mistral can return:
- `string` → historical path. `Answer = content`, `ReasonContent = ""`.
Preserves behavior for every non-reasoning model (`mistral-large-*`,
`mistral-small-*`, `ministral-*`, `codestral-*`, `pixtral-*`,
`open-mistral-nemo`).
- `[]interface{}` → walk the parts. Concatenate every `{"type":"text",
"text":...}` part into `Answer`; concatenate the inner text inside every
`{"type":"thinking", "thinking":[...]}` part into `ReasonContent`.
- `ChatWithMessages` now calls the helper instead of doing the raw
`.(string)` cast.
- Unknown part types are **skipped, not failed**. Mistral has been
adding new content variants quickly (audio chunks, citations, etc.);
this driver should not 500 every call when a new part type appears.
- `conf/models/mistral.json`: add `magistral-medium-latest` and
`magistral-small-latest`. Both are visible in `/v1/models` today.
No interface change. No factory change. No new dependencies.
### How was this tested?
**Unit tests** — 5 new tests in `internal/entity/models/mistral_test.go`
on top of the 27 already shipped via #14807:
- `TestMistralChatHandlesStringContent` — regression net for the
historical path
- `TestMistralChatExtractsReasoningFromStructuredContent` — the fixture
body is a trimmed copy of the actual `magistral-medium-latest` response
captured above; asserts both `Answer` and `ReasonContent` are populated
correctly
- `TestMistralChatHandlesStructuredContentWithoutThinking` —
`magistral-*` with a trivial answer returns a structured shape that has
only a `text` part; `ReasonContent` must stay empty
- `TestMistralChatIgnoresUnknownContentPartTypes` — `audio_url` and
`future_part_type` parts are skipped, `text` parts still flow through
- `TestExtractMistralContent` — table-driven unit coverage of the helper
for string, empty string, nil, empty array, text-only, thinking+text,
unsupported root type
```
$ go test -vet=off -run "TestMistral|TestExtractMistralContent" -count=1 -v ./internal/entity/models/...
=== RUN TestMistralChatHandlesStringContent
--- PASS: TestMistralChatHandlesStringContent (0.00s)
=== RUN TestMistralChatExtractsReasoningFromStructuredContent
--- PASS: TestMistralChatExtractsReasoningFromStructuredContent (0.00s)
=== RUN TestMistralChatHandlesStructuredContentWithoutThinking
--- PASS: TestMistralChatHandlesStructuredContentWithoutThinking (0.00s)
=== RUN TestMistralChatIgnoresUnknownContentPartTypes
--- PASS: TestMistralChatIgnoresUnknownContentPartTypes (0.00s)
=== RUN TestExtractMistralContent
=== RUN TestExtractMistralContent/plain_string
=== RUN TestExtractMistralContent/empty_string
=== RUN TestExtractMistralContent/nil
=== RUN TestExtractMistralContent/empty_array
=== RUN TestExtractMistralContent/text_only
=== RUN TestExtractMistralContent/thinking_then_text
=== RUN TestExtractMistralContent/unknown_root_type
--- PASS: TestExtractMistralContent (0.00s)
PASS
ok ragflow/internal/entity/models 0.046s
```
All 32 Mistral tests pass on go 1.25. `go build
./internal/entity/models/...` exits 0.
**Live integration test** — driver exercised against `api.mistral.ai/v1`
with the patched code:
```
=== RUN TestMistralMagistralSmoke
[OK] "magistral-small-latest" present upstream
[OK] "magistral-medium-latest" present upstream
[OK trivial] Answer="7" ReasonContent=""
[OK reasoning] Answer len=797 head="To determine when the two trains meet, we can follow these steps:\n\n1. **Identify..."
ReasonContent len=1069 head="Okay, let's see. There are two trains, one going 60 mph and the other going 90 mph. They're moving towards each other, s..."
MAGISTRAL SMOKE PASSED
--- PASS: TestMistralMagistralSmoke (18.09s)
PASS
ok ragflow/internal/entity/models 18.112s
```
What the live run proves on the wire:
- `magistral-small-latest` with a trivial prompt still uses the
string-content shape; the regression-net path is exercised against the
real server, not just the mock.
- `magistral-medium-latest` with a reasoning prompt uses the
structured-array shape; the new code path extracts a 1069-character
reasoning trace into `ChatResponse.ReasonContent` and a 797-character
visible answer into `ChatResponse.Answer`. Before this fix, the same
call returned `"invalid content format"` and the caller saw nothing.
The smoke-test file itself is not committed (live tests live outside the
PR diff, same convention used for prior provider PRs).
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
## Problem
The Go server build pipeline (`build.sh` + CMake + CGO bindings) was
tested on Ubuntu only. On macOS arm64 with Homebrew it fails in five
orthogonal places. None of these require platform-specific code paths —
the same source builds on both Linux and Darwin after these fixes.
## Reproduction (before)
```
$ uname -a
Darwin … 25.4.0 arm64
$ brew install cmake pcre2 simde
$ bash build.sh
…
error: 'simde/x86/sse4.1.h' file not found
error: implicit instantiation of undefined template 'std::basic_istringstream<char>'
error: no matching function for call to 'Join'
…
clang: error: no such file or directory: '/usr/local/lib/libpcre2-8.a'
```
## Fix (5 small, orthogonal changes)
### 1. `internal/cpp/CMakeLists.txt` — find Homebrew + libpcre2-8
portably
- Detect Apple platforms via `if(APPLE)`, call `brew --prefix` once, add
`${HOMEBREW_PREFIX}/include` and `${HOMEBREW_PREFIX}/lib`. No effect on
Linux.
- Replace the literal `libpcre2-8.a` link token (which only the Linux
linker finds in `/usr/local/lib` by default) with
`find_library(PCRE2_LIB NAMES pcre2-8 REQUIRED)`. Works on
`/usr/lib/x86_64-linux-gnu` (Debian/Ubuntu), `/usr/local/lib` (Intel Mac
& legacy Linux), `/opt/homebrew/lib` (Apple Silicon).
### 2. `internal/cpp/wordnet_lemmatizer.cpp` +
`internal/cpp/rag_analyzer.cpp` — explicit `#include <sstream>`
libstdc++ (Linux) pulls `<sstream>` in transitively via `<fstream>`;
libc++ (Apple Clang) doesn't, so the existing `std::istringstream` /
`std::ostringstream` uses fail to compile on macOS. One-line include in
each file.
### 3. `internal/cpp/rag_analyzer.cpp` — `Join` template overload fix
`Join(tokens, start, tokens.size(), delim)` at line 146 passes `size_t`
to an `int` parameter. C++23 strict mode in Apple Clang refuses the
implicit narrowing and reports the 4-arg overload as a substitution
failure, leaving the call ambiguous between the 3-arg and 4-arg
templates. Fix: explicit `static_cast<int>(tokens.size())`. Behaviour
identical on libstdc++ — the narrowing was always intentional.
### 4. `internal/binding/rag_analyzer.go` — split darwin CGO LDFLAGS
The existing `#cgo darwin LDFLAGS: ... /usr/local/lib/libpcre2-8.a` only
matches Intel Macs. Apple Silicon Homebrew installs to `/opt/homebrew`.
Split into `darwin,arm64` and `darwin,amd64` build constraints with the
right absolute path on each.
### 5. `build.sh` — accept Homebrew path in the pcre2 sanity check
The sanity check looked at two Linux paths only and then fell through to
`sudo apt -y install libpcre2-dev` on failure. Added
`/opt/homebrew/lib/libpcre2-8.a`, and on Darwin failure now exits
cleanly with the right `brew install pcre2` hint instead of trying
`apt`.
## Verified
- `bash build.sh` now completes on macOS arm64 (Apple Silicon, brew 4.x,
cmake 4.x, Apple Clang 17, Go 1.25, pcre2 10.x, simde 0.8.x).
- Produced binaries: `bin/server_main`, `bin/admin_server`,
`bin/ragflow_cli`.
- `bin/server_main` boots, connects MySQL, runs migrations, loads the 64
model provider configs cleanly.
- Still builds on Linux — the CMake additions are inside an `if(APPLE)`
guard, the `find_library` call matches Linux paths too, the build.sh
check still tries `apt` when not on Darwin.
## Out of scope
The Go server itself currently fails at runtime when not pointing at
Elasticsearch (`Failed to initialize doc engine: failed to ping
Elasticsearch`), but that's the placeholder Infinity engine documented
in `internal/engine/README.md` — unrelated to this build patchset.
---
Happy to split this into smaller PRs if you'd prefer (one per file). The
five changes are independent.
## What
- Add Perplexity as a chat and embedding provider backed by its
OpenAI-compatible `/chat/completions` and `/v1/embeddings` APIs
- Register Perplexity in the Go model factory and provider config
- Support non-streaming chat, SSE streaming chat, embeddings, model
listing, and connection checks
Refs #14736
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
- Adds an `Astraflow` Go driver so the new API server can route
Astraflow (UCloud ModelVerse) chat instances, matching the existing
Python `AstraflowChat` (`rag/llm/chat_model.py:1237`). Follows the same
SaaS-driver shape used for Avian, Novita, TogetherAI, Replicate,
DeepInfra, Upstage, and LongCat.
Closes#15062
---------
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
Closes#15044.
Avian was listed unchecked in the Go-rewrite tracker #14736 and already
had an llm_factories.json entry with 4 preconfigured chat models
(deepseek-v3.2, kimi-k2.5, glm-5, minimax-m2.5), but the Go API server
had no driver to route them. The Python side has supported Avian at
rag/llm/chat_model.py:1220 (AvianChat) via the LiteLLM openai/ provider
with default base https://api.avian.io/v1.
Co-authored-by: Jin Hai <haijin.chn@gmail.com>
