Files
ragflow/internal/tokenizer/bpe_loader.go
Jack e997fd655a fix(tokenizer): load cl100k BPE table from disk instead of failing silently offline (#17712)
## Summary

RAGFlow's Go tokenizer silently returned **0 tokens for every string**
whenever the `cl100k_base` BPE table could not be loaded — which is the
normal case for an offline/air-gapped Go server. This PR makes the
loader resolve the table from disk (where RAGFlow actually ships it) and
fail loudly when it is genuinely missing.

## Root cause

`tiktoken-go`'s stock loader downloads the encoding table over HTTP and
caches it under `TIKTOKEN_CACHE_DIR`. That does not work for RAGFlow:

- `TIKTOKEN_CACHE_DIR` is exported **only inside the Python process**
(`common/token_utils.py`). `docker/entrypoint.sh` launches the Go binary
(`bin/ragflow_server`) from a shell, so the Go process never inherits
the variable.
- The Dockerfile *does* ship the table (under its sha1 name in the
working directory), but nothing told the Go side to look there.
- Reaching `openaipublic.blob.core.windows.net` at runtime is not an
option for air-gapped installs, and is unreliable where that host is
blocked.

The failure was **silent**: `NumTokensFromString` returns `0` when the
encoder fails to build, and a `sync.Once` memoizes that error for the
process lifetime. Every token count became `0`, so chunk merging never
crossed its token budget and an entire document collapsed into a single
chunk. Python has no such failure mode because its encoder is built at
import time (a missing table aborts startup instead of degrading).

## Fix

Register a local-only `BpeLoader` via `tiktoken.SetBpeLoader`
(`internal/tokenizer/bpe_loader.go`) that resolves the table from disk
**only**, in priority order:

1. `TIKTOKEN_CACHE_DIR` / `DATA_GYM_CACHE_DIR` (honored so operators who
already configured one keep working).
2. The working directory, the executable's directory, and all of their
ancestors — matching the Dockerfile layout (table under its sha1 name in
the install root).
3. A `ragflow_deps/<basename>` checkout produced by
`ragflow_deps/download_deps.py`.

It **never performs network I/O**. When nothing is found it returns an
error listing every path it tried (pointing at `download_deps.py` or
`TIKTOKEN_CACHE_DIR`), so a genuinely missing table fails loudly instead
of degrading to zero.

## Test plan

- `internal/tokenizer/bpe_loader_test.go` (unit tier, runs under `bash
build.sh --test ./internal/tokenizer/...`):
- Loader reads from `TIKTOKEN_CACHE_DIR`, `DATA_GYM_CACHE_DIR`, the
sha1-named file in the working dir, and the bundled `ragflow_deps/`
name.
  - Explicit cache dir wins over the bundled vocab.
  - A malformed table is reported as an error rather than skipped.
- A genuinely missing table reports the candidates it tried (no network
attempt).
- `NumTokensFromString` matches Python-derived anchors (`""`→0,
`"hello"`→1, `"hello world"`→2, `"hello, world!"`→4, `"世界"`→3, `"Hello
世界 🌍"`→8, `"RAGFlow"`→3).

## Notes

- `.github/workflows/tests.yml` currently excludes `internal/tokenizer`
from `go test`, so these tests do not run in CI. The tokenizer fix is
exercised in CI indirectly via the chunker package once a
token-count-sensitive parity case lands (tracked separately). Consider
including `internal/tokenizer` in CI as a follow-up.
- Supported deployments already ship the table (`download_deps.py` →
`ragflow_deps/cl100k_base.tiktoken`; Dockerfile → `<sha1>` in cwd), so
no `ENV` change is required for the fix to take effect. Setting `ENV
TIKTOKEN_CACHE_DIR` in the Dockerfile remains a cheap
belt-and-suspenders hardening that can be done separately.

🤖 Generated with [CodeBuddy Code](https://cnb.cool/codebuddy)

---------

Co-authored-by: CodeBuddy <noreply@codebuddy.ai>
Co-authored-by: CodeBuddy Code <noreply@cnb.cool>
Co-authored-by: CodeBuddy <noreply@tencent.com>
2026-08-03 19:03:08 +08:00

217 lines
7.4 KiB
Go

//
// Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
package tokenizer
// Offline BPE table loading for tiktoken.
//
// RAGFlow ships the cl100k_base table on disk (Dockerfile drops it into the
// working directory under its sha1 name; download_deps.py writes it to
// ragflow_deps/). tiktoken-go's stock loader instead downloads it over HTTP and
// relies on TIKTOKEN_CACHE_DIR, which the Go server never inherits, so a
// missing table degrades every token count to 0. This loader resolves the
// table from disk only: it performs no network I/O, and when nothing is found
// it reports every path it tried.
import (
"crypto/sha1"
"encoding/base64"
"fmt"
"os"
"path"
"path/filepath"
"strconv"
"strings"
"ragflow/internal/common"
"github.com/pkoukk/tiktoken-go"
)
func init() {
tiktoken.SetBpeLoader(localBpeLoader{})
}
// expectedBpeHashes maps a tiktoken table URL to the SHA-1 of its canonical
// on-disk content. We only ship cl100k_base today; entries here let the loader
// reject a corrupt or tampered file instead of trusting it. Unknown URLs are
// loaded without a digest check (defense-in-depth, not a hard gate).
//
// NOTE: this is the digest of the file *contents*, not the tiktoken cache
// filename. tiktoken-go names its cached file by sha1(bpeURL)
// (223921b76ee99bde995b7ff738513eef100fb51d18c93597a113bcffe865b2a7 for
// cl100k_base); that value identifies the path, while the value below verifies
// the bytes we actually load. Compute it from the table shipped by
// ragflow_deps/download_deps.py: `sha1sum cl100k_base.tiktoken`.
var expectedBpeHashes = map[string]string{
"https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken": "6494e42d5aad2bbb441ea9793af9e7db335c8d9c",
}
// localBpeLoader resolves tiktoken BPE tables from the local filesystem.
type localBpeLoader struct{}
// LoadTiktokenBpe implements tiktoken.BpeLoader.
//
// bpeURL is the upstream table URL that tiktoken-go would otherwise download;
// here it serves only to derive the file names to look for.
func (localBpeLoader) LoadTiktokenBpe(bpeURL string) (map[string]int, error) {
candidates := bpeCandidatePaths(bpeURL)
for _, candidate := range candidates {
contents, err := os.ReadFile(candidate)
if err != nil {
// Only a missing candidate is skippable; a permission or I/O
// failure must not be masked as "not found".
if os.IsNotExist(err) {
continue
}
return nil, fmt.Errorf("reading BPE table %s: %w", candidate, err)
}
// Integrity check: for tables we ship, a digest mismatch means the
// file is corrupt or tampered with. Refuse to load it rather than
// skipping to the next candidate — a different candidate holds the
// same (wrong) content, and masking the failure would defeat the
// check. This mirrors the malformed-table path just below.
if want, ok := expectedBpeHashes[bpeURL]; ok {
if got := fmt.Sprintf("%x", sha1.Sum(contents)); got != want {
return nil, fmt.Errorf("BPE table %s digest mismatch (got %s, want %s); refusing to load a corrupt or tampered file", candidate, got, want)
}
}
ranks, err := parseBpeTable(contents)
if err != nil {
// A file that exists but does not parse is a corrupt download or a
// name collision. Continuing to the next candidate would mask it.
return nil, fmt.Errorf("BPE table %s is malformed: %w", candidate, err)
}
return ranks, nil
}
err := fmt.Errorf(
"no local BPE table for %s; run `uv run ragflow_deps/download_deps.py` or set TIKTOKEN_CACHE_DIR to the directory holding the table; tried: %s",
bpeURL, strings.Join(candidates, ", "))
// Logged as well as returned: tiktoken-go propagates this to GetEncoding,
// whose error NumTokensFromString discards to keep returning 0.
common.Error("cl100k BPE table not found; every token count will be 0", err)
return nil, err
}
// bpeCandidatePaths lists, in priority order, every local path that may hold
// the table for bpeURL.
//
// Explicit configuration wins, then the directories RAGFlow actually ships the
// table in. Both the working directory and the executable's directory are
// walked upwards: the server runs with the working directory set to the
// installation root, while `go test` runs from a package subdirectory.
func bpeCandidatePaths(bpeURL string) []string {
cacheName := fmt.Sprintf("%x", sha1.Sum([]byte(bpeURL)))
// download_deps.py stores the table under the URL's own basename.
bundledName := path.Base(bpeURL)
var paths []string
seen := make(map[string]struct{})
add := func(p string) {
if _, dup := seen[p]; dup {
return
}
seen[p] = struct{}{}
paths = append(paths, p)
}
// Honour both variables tiktoken-go itself reads, so an operator who has
// already configured one keeps working.
for _, env := range []string{"TIKTOKEN_CACHE_DIR", "DATA_GYM_CACHE_DIR"} {
if dir := strings.TrimSpace(os.Getenv(env)); dir != "" {
add(filepath.Join(dir, cacheName))
}
}
for _, root := range searchRoots() {
// Same layout the Dockerfile creates: the table sits in the
// installation root under its sha1 name.
add(filepath.Join(root, cacheName))
// download_deps.py writes the table into ragflow_deps/ under its
// download name; a developer checkout that has run it but never
// started the Python side only has this copy.
add(filepath.Join(root, "ragflow_deps", bundledName))
}
return paths
}
// searchRoots returns the working directory and the executable's directory
// together with all of their ancestors.
func searchRoots() []string {
var roots []string
seen := make(map[string]struct{})
for _, start := range startingDirs() {
for dir := start; ; {
if _, dup := seen[dir]; !dup {
seen[dir] = struct{}{}
roots = append(roots, dir)
}
parent := filepath.Dir(dir)
if parent == dir {
break
}
dir = parent
}
}
return roots
}
func startingDirs() []string {
var dirs []string
if wd, err := os.Getwd(); err == nil {
dirs = append(dirs, wd)
}
if exe, err := os.Executable(); err == nil {
if resolved, err := filepath.EvalSymlinks(exe); err == nil {
exe = resolved
}
dirs = append(dirs, filepath.Dir(exe))
}
return dirs
}
// parseBpeTable decodes tiktoken's on-disk format: one
// "<base64 token> <rank>" pair per line.
func parseBpeTable(contents []byte) (map[string]int, error) {
ranks := make(map[string]int)
for i, line := range strings.Split(string(contents), "\n") {
line = strings.TrimRight(line, "\r")
if line == "" {
continue
}
token, rank, ok := strings.Cut(line, " ")
if !ok {
return nil, fmt.Errorf("line %d: expected \"<token> <rank>\"", i+1)
}
decoded, err := base64.StdEncoding.DecodeString(token)
if err != nil {
return nil, fmt.Errorf("line %d: %w", i+1, err)
}
value, err := strconv.Atoi(rank)
if err != nil {
return nil, fmt.Errorf("line %d: %w", i+1, err)
}
ranks[string(decoded)] = value
}
if len(ranks) == 0 {
return nil, fmt.Errorf("table is empty")
}
return ranks, nil
}