Some NVIDIA hosted models (e.g. meta/llama-3.2-11b-vision-instruct )
expose a full
endpoint URL per model that does not follow the normal base_url +
url_suffix
assembly. Previously the Go driver always called {base}/chat/completions
, so chat
requests for these vision models hit the wrong endpoint and failed.
This PR adds an optional per-model url field in conf/models/nvidia.json
. When
present, every NVIDIA driver request (chat, streaming chat, embedding,
rerank, model
listing) uses it directly; otherwise the standard assembly is unchanged.
Add a shared `dao.ResolveModelContentLength` that resolves a chat model's context window (`content_length`) from a `tenant_model` UUID or a composite `model@provider` reference, with an optional `driver + modelName` catalog fallback for the no-database path.
The agent LLM component and the ingestion Extractor component both need
to trim prompts to the model's context window before calling the
provider. Each previously did (or would do) this with its own copy of
the logic. This PR adds the shared primitive; follow-up PRs wire it into
the agent LLM component (#18092) and the ingestion Extractor/tagger
(#18095).
### Summary
`format_document_soup` tracks "am I inside a table" and "am I inside a
link" with sticky flags that are meant to be reset by `elif e.name ==
"/table"` and `elif e.name == "/a"`. BeautifulSoup's `.descendants` only
yields opening tags — a `Tag` named `/table` or `/a` never exists — so
both branches are dead code and neither flag is ever cleared.
Everything after the first `<table>` on a page is therefore formatted as
if it were still table content: paragraphs lose their newline, list
items lose their `- ` marker, headings lose their break, and the text is
glued onto the last table cell. Under
`HTML_BASED_CONNECTOR_TRANSFORM_LINKS_STRATEGY=markdown` the same bug
leaks a link's `href` into everything that follows it, including whole
subsequent paragraphs. The Confluence connector
(`confluence_connector.py:948`) goes through this path.
Real output for a Confluence-shaped page (heading, intro, spec table,
then the body) via the public `parse_html_page_basic`:
**Before**
```
prod us-east-1 Rollback procedure If the canary fails, run the rollback script immediately. Drain the load balancer Revert the deployment Escalate to the on-call rota if the rollback stalls. Do not skip the post-mortem.
```
**After**
```
prod us-east-1
Rollback procedure
If the canary fails, run the rollback script immediately.
- Drain the load balancer
- Revert the deployment
Escalate to [the on-call rota](http://oncall.example.com) if the rollback stalls.
Do not skip the post-mortem.
```
Every heading, paragraph and list marker after the table is lost, and
the whole body is indexed as one run-on line hanging off a table cell.
### Fix
Derive both scopes from each element's **ancestors** instead of from
flags that nothing can clear, and drop the two dead branches plus the
two that become redundant.
The scopes are resolved in one up-front pass into `id`-keyed maps
(`table_scope`, `href_scope`) and looked up in O(1) per element. Probing
per element with `find_parent` instead is O(depth) each, which measured
12–13× slower on table-heavy pages and up to 103× on deeply nested
markup; the map version costs a depth-independent 1.13–1.35× over
`main`. Numbers and method are in the round-2 comment below.
This also changes one adjacent behaviour worth calling out explicitly: a
link **inside** a table cell now renders as markdown, where before it
rendered as plain text. That previous behaviour was not by design — it
only held when no link preceded the table. With a link before the table,
`main` stamps the stale href onto every cell:
```
main: '[pre](http://STALE.com)\n\t[cellA](http://STALE.com)\t[cellB](http://STALE.com)'
branch: '[pre](http://STALE.com)\n\tcellA\tcellB'
```
Those cells are not links. Both symptoms are the same sticky-state bug,
so they are fixed together rather than left half-done.
### Testing
`test/unit_test/data_source/test_html_utils.py` is new —
`format_document_soup` had no test coverage. 11 tests: 8 fail on `main`
and pass on this branch, 3 are controls that pass on both (the table
itself still separates rows and cells, anchor text is still linkified,
the default `strip` strategy still strips).
Representative failures on `main`:
```
assert '\nAfter' in 'Before\n\tA\tB After'
assert '\n- item1' in 'Before\n\tA\tB item1 item2'
assert 'see [link](http://x.com) [ after](http://x.com)' == 'see [link](http://x.com) after'
assert '[next paragraph]' not in '[link](http://x.com)\n[next paragraph](http://x.com)'
```
Reverting each clause of the fix independently keeps the anchors honest:
reverting only the table clause fails exactly the 4 table tests and
leaves the link tests green; reverting only the link clause fails
exactly the 3 link tests and leaves the table tests green.
(`test_link_inside_a_table_cell_is_linkified` needs both clauses broken
to fail, so it appears in neither single-clause revert — it is covered
by the 8-fail run against `main`.)
Full `test/unit_test/data_source/` suite: **3 failed, 199 passed**, and
the failure set is byte-identical to clean `main` (**3 failed, 188
passed**) — the 3 are `TestSSRFValidation::*`, which resolve
`api.example.com` against real DNS and are unrelated to this change.
`ruff check` and `ruff format --check` are clean on both touched files.
---
This PR was drafted with AI assistance (Claude). I reviewed the change,
independently reproduced both symptoms against `main`, and take
responsibility for it.
### Summary
- Propagate the dataset language through Go DOCX, Markdown, PDF
figure-enhancement, and standalone-image vision paths.
- Explicitly render the shared figure prompt's `{{ language }}`
placeholder in Go.
- Use English when the dataset language is empty.
- Make the default standalone-image prompt request the dataset language
while preserving visible text in its original language.
- Add focused tests for caller propagation, language fallback, prompt
rendering, and prompt-cache isolation.
Stop flattening `<table>` into a single text blob. A `<table>` now emits:
1. an inlined `doc_type_kwd:"text"` item keeping the `<table>…</table>`
markup (row/column structure survives for embedding/retrieval/LLM rendering),
2. a structured `doc_type_kwd:"table"` / `ck_type:"table"` item appended
after the walk, consumed by the downstream chunker.
Port the wiki_incremental dataset-level merge and make its rewrite
barrier durable and concurrency-safe. Wiki pages merge replace-only; the
barrier persists a monotonic numeric generation, and a scheduler-backed
per-dataset lock closes the cross-process TOCTOU window. Adds the
Compiler Plan toggle (frontend) with Mode A grouping.
### Summary
Fixes#18107.
`editdistance==0.8.1` (the only recent release on PyPI) has no cp313
wheels for any platform. Since this project requires exactly Python
3.13, `uv`/`pip`/`poetry` fall back to building it from source (Cython),
which fails on Windows for anyone without a working C build toolchain —
that's the PEP 517 build error in the issue.
Swapped `editdistance` for `rapidfuzz`, which ships full cp313 wheels
(win32/win_amd64/win_arm64 included) and has no build-from-source step
on any of our target platforms. The only call site was
`EntityResolution.is_similarity` in `rag/graphrag/entity_resolution.py`,
using `editdistance.eval(a, b)` to get the unweighted Levenshtein
distance between two entity names.
`rapidfuzz.distance.Levenshtein.distance(a, b)` computes the same thing
(verified identical output on several string pairs) and is used as a
direct replacement.
### Summary
Refs #17885.
Mistral figure enrichment now receives the dataset language through the
production parsing path. `by_mistral_ocr` forwards `lang` to
`MistralParser.parse_pdf`; the parser stores the normalized language and
passes it to the figure-description prompt. Empty or missing values
still fall back to English.
### Summary
Brings both halves of the Tenki sandbox provider onto current SDKs and
removes `project_id`, which Tenki deleted from its API.
**Go:** `github.com/LuxorLabs/tenki-sdk-go/sandbox` `v0.5.2` → `v0.7.0`
(current latest).
**Python:** the provider's SDK was renamed on PyPI — `tenki-sandbox` is
frozen at 0.4.0 and everything from 0.5 ships as
[`tenki`](https://pypi.org/project/tenki/). The docs told operators to
`pip install tenki-sandbox`, which installs a stale SDK that no longer
matches this provider's expectations.
**`project_id` is gone.** Tenki removed project scoping from the sandbox
API in 0.5.x: `Client.create()` no longer accepts `project_id`, so the
current code path would raise `TypeError` against a current SDK. It was
also marked `required: True` in the config schema, so the Admin >
Sandbox Settings form asked for a value that no longer exists.
### What problem does this PR solve?
Incremental Wiki compilation could lose provenance for claim-light
entities, produce unstable page groups across embedding models, route
entities to unrelated pages, and assign topics without sufficient
page-level context. Document removals and page membership changes could
also leave stale Wiki state.
This PR:
- preserves source document and chunk provenance throughout entity
matching, reduction, page generation, and deletion;
- uses embeddings to retrieve candidates and the LLM to make final page
grouping and incremental routing decisions;
- batches embedding and LLM operations with bounded concurrency and
deterministic fallbacks;
- selects source-scoped topic candidates with embeddings before the page
LLM chooses the final topic;
- rebuilds Wiki state when the compilation mode or embedding model
changes;
- normalizes Wiki array fields returned by the API and retains entities
without relations in graph responses.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
## What
This pull request adds **MWS GPT Model Hub** as a built-in model
provider in RAGFlow.
The integration allows users to configure an MWS project endpoint and
token, discover the models available to that project, and use supported
MWS models for chat completion, embeddings, and reranking.
Co-authored-by: ilarionov_n <ilarionov_n@promis.ru>
Centralize the shared golden-doc + alignment helpers in `align_test.go` so the format-specific PRs (text&code, markdown golden, HTML) reuse one implementation instead of each carrying their own copy of the scaffolding.
Golden-parity test infrastructure for the **Go `TokenChunker` ↔ Python alignment**.
It runs the Go chunker over a committed case set (`testdata/parity/cases/`) and diffs each output against a captured Python golden (`testdata/parity/golden/`), honoring a `known_diffs.json` ratchet (`extra_fields` / `chunk_count` / `chunk_text`) so accepted divergences are tracked rather than silently widening.
