mirror of
https://github.com/mudler/LocalAI.git
synced 2026-06-18 21:58:58 -04:00
feat(pii): NER tier engine — privacy-filter.cpp backend + NER-centric PII filter (#10360)
Squashed feat/pii-ner-tier-engine rebased onto master (was 45 commits; see backup/pii-ner-tier-engine-prerebase). Net change: - privacy-filter.cpp: standalone GGML engine for the openai-privacy-filter PII/NER token classifier, wired as a LocalAI gRPC backend (CPU/CUDA/Vulkan). TokenClassify moves off the patched llama.cpp path onto this backend. - PII filter reworked to be NER-centric (encoder/NER detection tier scanning whole conversations as one document), with a recreated bounded restricted- regex secret-matching pattern detector tier alongside it (per-model pii_detection.builtins / .patterns + core/services/routing/piipattern). - Detection labelled by source (ner vs pattern); backend trace / confidence / debug observability; analyze/redact exposed as a synchronous API. - Instance-wide default detector policy + per-usecase default-on; request filtering extended to completions, embeddings, edits & Ollama. - React UI: NER-centric PII editor, detector-models table, pattern/builtins editor, middleware default-policy UI. - Gallery: privacy-filter-multilingual token-classify model + NER install filter; token_classify known_usecase; batch sized to context for NER models. privacy-filter backend registered in the backend gallery (cpu/vulkan/cuda-13 meta + image entries with a capabilities map) matching its CI matrix jobs, and an /import-model auto-detect importer (PrivacyFilterImporter, narrow privacy-filter GGUF detection) replacing the prior pref-only registration. Reconciled against master's independent evolution: - Dropped master's PIIPatternOverrides feature (global-pattern runtime overrides + /api/pii/patterns API + runtime_settings.json persistence). The per-model NER + pattern-detector design supersedes it; it was built on the global redactor pattern set this branch replaced. - Reverted the llama.cpp Score carry-patch (0006-server-task-type-score): removed the patch and restored master's grpc-server.cpp Score RPC (direct llama_decode, slot-loop bypass) and LLAMA_VERSION pin, plus master's model_config validation forbidding score + chat/completion/embeddings on llama-cpp. token_classify is unaffected (it runs on the privacy-filter backend, not llama-cpp). Assisted-by: Claude:claude-opus-4-8 [Claude Code] Signed-off-by: Richard Palethorpe <io@richiejp.com>
This commit is contained in:
committed by
GitHub
parent
c133ca39dc
commit
3fa7b2955c
@@ -184,23 +184,11 @@ func nonStreamingOpenAIToolCallScript() (status int, body string, contentType st
|
||||
return 200, `{"id":"chatcmpl-y","choices":[{"index":0,"message":{"role":"assistant","content":"","tool_calls":[{"id":"call_lookup","type":"function","function":{"name":"lookup","arguments":"{\"q\":\"clouds\"}"}}]},"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":12,"completion_tokens":7,"total_tokens":19}}`, "application/json"
|
||||
}
|
||||
|
||||
// emailLeakOpenAIScript returns a non-streaming response containing an
|
||||
// email address. The streaming PII filter doesn't apply to buffered
|
||||
// responses, but the response is JSON the client receives unchanged —
|
||||
// used to verify the wire path without PII assertions. The streaming
|
||||
// PII variant uses an SSE response.
|
||||
func emailLeakOpenAIStreamingScript() (status int, body string, contentType string) {
|
||||
frames := []string{
|
||||
`{"choices":[{"index":0,"delta":{"content":"contact alice@"}}]}`,
|
||||
`{"choices":[{"index":0,"delta":{"content":"example.com please"}}]}`,
|
||||
`{"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}`,
|
||||
}
|
||||
var b strings.Builder
|
||||
for _, f := range frames {
|
||||
b.WriteString("data: ")
|
||||
b.WriteString(f)
|
||||
b.WriteString("\n\n")
|
||||
}
|
||||
b.WriteString("data: [DONE]\n\n")
|
||||
return 200, b.String(), "text/event-stream"
|
||||
}
|
||||
// leakedAnthropicKey is a synthetic Anthropic API key (sk-ant-… shape) used by
|
||||
// the translate-mode PII test: the client sends it inside a user message and
|
||||
// the request-side pattern detector (anthropic_api_key builtin, which matches
|
||||
// sk-ant-[A-Za-z0-9_-]{20,}) must catch it and block the request before the
|
||||
// cloud-proxy forwards anything upstream. An email is deliberately not used —
|
||||
// the pattern tier rejects open-ended shapes (that's the NER tier's job, and
|
||||
// the NER model can't run in this hermetic suite).
|
||||
const leakedAnthropicKey = "sk-ant-api03-ABCDEFGHIJKLMNOPQRSTUVWXYZ012345"
|
||||
|
||||
@@ -225,36 +225,31 @@ var _ = Describe("Cloud-proxy backend E2E", func() {
|
||||
})
|
||||
|
||||
Context("Translate mode + PII filter", func() {
|
||||
It("applies the streaming PII filter to translate-mode content", func() {
|
||||
// Default PII config redacts email addresses. Split the
|
||||
// email across two SSE deltas so the filter has to buffer
|
||||
// the partial match — proves the streaming filter is wired
|
||||
// up in translate mode, not just passthrough.
|
||||
It("blocks request-side secrets before they reach the upstream", func() {
|
||||
// Our NER/pattern PII tier is request-side by design: it scans
|
||||
// inbound content, not upstream responses. cp-translate-openai
|
||||
// wires the in-process pattern detector (e2e-secret-filter, block
|
||||
// action), so a leaked secret in the user's message must trip the
|
||||
// request-side filter and the request must be rejected before the
|
||||
// cloud-proxy ever forwards it to the provider.
|
||||
var upstreamHit bool
|
||||
cpOpenAIUpstream.SetScript(func([]byte) (int, string, string) {
|
||||
return emailLeakOpenAIStreamingScript()
|
||||
upstreamHit = true
|
||||
return 200, `{"id":"x","choices":[{"index":0,"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}`, "application/json"
|
||||
})
|
||||
|
||||
cp := openai.NewClient(option.WithBaseURL(apiURL))
|
||||
stream := cp.Chat.Completions.NewStreaming(context.TODO(), openai.ChatCompletionNewParams{
|
||||
_, err := cp.Chat.Completions.New(context.TODO(), openai.ChatCompletionNewParams{
|
||||
Model: "cp-translate-openai",
|
||||
Messages: []openai.ChatCompletionMessageParamUnion{
|
||||
openai.UserMessage("share contact info"),
|
||||
openai.UserMessage("my key is " + leakedAnthropicKey + " please keep it"),
|
||||
},
|
||||
})
|
||||
|
||||
var assembled strings.Builder
|
||||
for stream.Next() {
|
||||
for _, ch := range stream.Current().Choices {
|
||||
assembled.WriteString(ch.Delta.Content)
|
||||
}
|
||||
}
|
||||
Expect(stream.Err()).NotTo(HaveOccurred())
|
||||
out := assembled.String()
|
||||
// If PII is wired up, the email is redacted before reaching
|
||||
// the client. If not, "alice@example.com" leaks through.
|
||||
// This is the lock-in test for gap #3.
|
||||
Expect(out).NotTo(ContainSubstring("alice@example.com"),
|
||||
"email leaked through translate-mode stream — PII filter not applied")
|
||||
// Lock-in: request-side PII fires in translate mode, the request is
|
||||
// rejected by content policy, and the secret never leaves the box.
|
||||
Expect(err).To(HaveOccurred(), "request carrying a leaked secret must be rejected")
|
||||
Expect(err.Error()).To(ContainSubstring("pii_blocked"))
|
||||
Expect(upstreamHit).To(BeFalse(), "blocked request must not reach the upstream provider")
|
||||
})
|
||||
})
|
||||
})
|
||||
|
||||
@@ -458,6 +458,27 @@ var _ = BeforeSuite(func() {
|
||||
"upstream_url": cpOpenAIUpstream.URL() + "/v1/chat/completions",
|
||||
"api_key_env": "CLOUD_PROXY_E2E_OPENAI_KEY",
|
||||
},
|
||||
// Wire the in-process pattern detector so the streaming PII
|
||||
// filter has something to redact in translate mode. The NER
|
||||
// tier needs the privacy-filter model (unavailable in this
|
||||
// hermetic suite); the pattern tier runs in-process with no
|
||||
// backend load, so it's what exercises the streaming plumbing.
|
||||
"pii": map[string]any{
|
||||
"detectors": []string{"e2e-secret-filter"},
|
||||
},
|
||||
},
|
||||
{
|
||||
// In-process pattern detector: backend "pattern" is resolved by
|
||||
// the PII NER resolver directly from this config (no backend is
|
||||
// ever loaded). It matches high-entropy anchored secrets;
|
||||
// cp-translate-openai references it above via pii.detectors.
|
||||
"name": "e2e-secret-filter",
|
||||
"backend": "pattern",
|
||||
"known_usecases": []string{"token_classify"},
|
||||
"pii_detection": map[string]any{
|
||||
"default_action": "block",
|
||||
"builtins": []string{"anthropic_api_key", "openai_api_key"},
|
||||
},
|
||||
},
|
||||
{
|
||||
"name": "cp-translate-anthropic",
|
||||
|
||||
Reference in New Issue
Block a user