feat: implement rag answerability gate graph
This commit is contained in:
mlogclub
@@ -0,0 +1,233 @@
|
||||
package executor
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"strings"
|
||||
"testing"
|
||||
|
||||
"cs-agent/internal/ai/rag"
|
||||
"cs-agent/internal/ai/runtime/internal/impl/callbacks"
|
||||
"cs-agent/internal/ai/runtime/internal/impl/retrievers"
|
||||
"cs-agent/internal/models"
|
||||
|
||||
"github.com/cloudwego/eino/components/model"
|
||||
"github.com/cloudwego/eino/schema"
|
||||
)
|
||||
|
||||
func TestParseAnswerabilityDecisionRejectsMalformedJSON(t *testing.T) {
|
||||
_, err := parseAnswerabilityDecision(`{"answerable": true`)
|
||||
|
||||
if err == nil {
|
||||
t.Fatal("expected malformed JSON to be rejected")
|
||||
}
|
||||
}
|
||||
|
||||
func TestParseAnswerabilityDecisionRejectsAnswerableWithoutSupportingChunkIDs(t *testing.T) {
|
||||
_, err := parseAnswerabilityDecision(`{"answerable": true, "reason": "directly supported"}`)
|
||||
|
||||
if err == nil {
|
||||
t.Fatal("expected answerable decision without supporting chunks to be rejected")
|
||||
}
|
||||
}
|
||||
|
||||
func TestParseAnswerabilityDecisionAcceptsAnswerableWithSupportingChunkIDs(t *testing.T) {
|
||||
got, err := parseAnswerabilityDecision("```json\n{\"answerable\": true, \"reason\": \"directly supported\", \"supportingChunkIds\": [\" chunk-1 \", \"chunk-2\"]}\n```")
|
||||
if err != nil {
|
||||
t.Fatalf("parse decision failed: %v", err)
|
||||
}
|
||||
|
||||
if !got.Answerable {
|
||||
t.Fatal("expected answerable decision")
|
||||
}
|
||||
if got.SupportingChunkIDs[0] != "chunk-1" || got.SupportingChunkIDs[1] != "chunk-2" {
|
||||
t.Fatalf("unexpected supporting chunks: %#v", got.SupportingChunkIDs)
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnowledgeAnswerabilityGateEvaluateFallsBackOnRetrieverError(t *testing.T) {
|
||||
collector := callbacks.NewRuntimeTraceCollector()
|
||||
gate := newTestKnowledgeAnswerabilityGate(&fakeKnowledgeContextRetriever{
|
||||
knowledgeBaseIDs: []int64{1},
|
||||
err: errors.New("vector store unavailable"),
|
||||
}, nil)
|
||||
|
||||
state, err := gate.Evaluate(context.Background(), answerabilityGateInput{
|
||||
Request: newAnswerabilityGateRunInput("是否支持退款？", "1"),
|
||||
Summary: &RunResult{},
|
||||
Collector: collector,
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatalf("Evaluate returned error: %v", err)
|
||||
}
|
||||
|
||||
if !strings.Contains(state.FallbackReply, "建议你联系人工客服进一步确认。") {
|
||||
t.Fatalf("expected human-support fallback, got %q", state.FallbackReply)
|
||||
}
|
||||
if collector.Data.Answerability.Status != answerabilityStatusUnanswerable {
|
||||
t.Fatalf("unexpected answerability status: %q", collector.Data.Answerability.Status)
|
||||
}
|
||||
if collector.Data.Answerability.Reason != "knowledge retrieval failed" {
|
||||
t.Fatalf("unexpected reason: %q", collector.Data.Answerability.Reason)
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnowledgeAnswerabilityGateEvaluateSkipsWhenNoKnowledgeConfigured(t *testing.T) {
|
||||
collector := callbacks.NewRuntimeTraceCollector()
|
||||
gate := newTestKnowledgeAnswerabilityGate(&fakeKnowledgeContextRetriever{}, nil)
|
||||
|
||||
state, err := gate.Evaluate(context.Background(), answerabilityGateInput{
|
||||
Request: newAnswerabilityGateRunInput("是否支持退款？", ""),
|
||||
Collector: collector,
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatalf("Evaluate returned error: %v", err)
|
||||
}
|
||||
|
||||
if !state.SkipGate {
|
||||
t.Fatal("expected gate to skip without knowledge")
|
||||
}
|
||||
if state.FallbackReply != "" {
|
||||
t.Fatalf("expected no fallback when gate skips, got %q", state.FallbackReply)
|
||||
}
|
||||
if collector.Data.Answerability.Status != answerabilityStatusSkipped {
|
||||
t.Fatalf("unexpected answerability status: %q", collector.Data.Answerability.Status)
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnowledgeAnswerabilityGateEvaluateFallsBackOnGrayZoneUnanswerableDecision(t *testing.T) {
|
||||
collector := callbacks.NewRuntimeTraceCollector()
|
||||
gate := newTestKnowledgeAnswerabilityGate(newAnswerabilityRetrieverWithHit(), &fakeAnswerabilityChatModel{
|
||||
response: `{"answerable": false, "reason": "retrieved snippets mention refunds but not the requested condition", "missingInfo": ["refund condition"]}`,
|
||||
})
|
||||
|
||||
state, err := gate.Evaluate(context.Background(), answerabilityGateInput{
|
||||
Request: newAnswerabilityGateRunInput("满足什么条件可以退款？", "1"),
|
||||
Collector: collector,
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatalf("Evaluate returned error: %v", err)
|
||||
}
|
||||
|
||||
if !strings.Contains(state.FallbackReply, "建议你联系人工客服进一步确认。") {
|
||||
t.Fatalf("expected human-support fallback, got %q", state.FallbackReply)
|
||||
}
|
||||
if collector.Data.Answerability.Status != answerabilityStatusUnanswerable {
|
||||
t.Fatalf("unexpected answerability status: %q", collector.Data.Answerability.Status)
|
||||
}
|
||||
if collector.Data.Answerability.MissingInfo[0] != "refund condition" {
|
||||
t.Fatalf("unexpected missing info: %#v", collector.Data.Answerability.MissingInfo)
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnowledgeAnswerabilityGateEvaluateAllowsAnswerableDecisionAndProducesKnowledgeInstruction(t *testing.T) {
|
||||
collector := callbacks.NewRuntimeTraceCollector()
|
||||
summary := &RunResult{}
|
||||
chatModel := &fakeAnswerabilityChatModel{
|
||||
response: `{"answerable": true, "reason": "refund condition is directly supported", "supportingChunkIds": ["101"]}`,
|
||||
}
|
||||
gate := newTestKnowledgeAnswerabilityGate(newAnswerabilityRetrieverWithHit(), chatModel)
|
||||
|
||||
state, err := gate.Evaluate(context.Background(), answerabilityGateInput{
|
||||
Request: newAnswerabilityGateRunInput("满足什么条件可以退款？", "1"),
|
||||
Summary: summary,
|
||||
Collector: collector,
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatalf("Evaluate returned error: %v", err)
|
||||
}
|
||||
|
||||
if state.FallbackReply != "" {
|
||||
t.Fatalf("expected answerable gate to allow, got fallback %q", state.FallbackReply)
|
||||
}
|
||||
if len(state.Decision.Instructions) != 1 {
|
||||
t.Fatalf("expected one knowledge instruction, got %d", len(state.Decision.Instructions))
|
||||
}
|
||||
if !strings.Contains(state.Decision.Instructions[0].Content, "知识库回答约束") {
|
||||
t.Fatalf("unexpected instruction: %q", state.Decision.Instructions[0].Content)
|
||||
}
|
||||
if summary.RetrieverCount != 1 {
|
||||
t.Fatalf("expected retriever count 1, got %d", summary.RetrieverCount)
|
||||
}
|
||||
if collector.Data.Answerability.Status != answerabilityStatusAnswerable {
|
||||
t.Fatalf("unexpected answerability status: %q", collector.Data.Answerability.Status)
|
||||
}
|
||||
if collector.Data.Answerability.SupportingChunkIDs[0] != "101" {
|
||||
t.Fatalf("unexpected supporting chunks: %#v", collector.Data.Answerability.SupportingChunkIDs)
|
||||
}
|
||||
if len(chatModel.input) == 0 || !strings.Contains(chatModel.input[len(chatModel.input)-1].Content, "chunkId: 101") {
|
||||
t.Fatalf("expected grader prompt to include chunk id, got %#v", chatModel.input)
|
||||
}
|
||||
}
|
||||
|
||||
func newTestKnowledgeAnswerabilityGate(retriever knowledgeContextRetriever, chatModel model.BaseChatModel) *KnowledgeAnswerabilityGate {
|
||||
return &KnowledgeAnswerabilityGate{
|
||||
newRetriever: func(aiAgent models.AIAgent) knowledgeContextRetriever {
|
||||
return retriever
|
||||
},
|
||||
newChatModel: func(ctx context.Context, aiConfig models.AIConfig) (model.BaseChatModel, error) {
|
||||
return chatModel, nil
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
func newAnswerabilityGateRunInput(question string, knowledgeIDs string) RunInput {
|
||||
return RunInput{
|
||||
UserMessage: models.Message{Content: question},
|
||||
AIAgent: models.AIAgent{
|
||||
KnowledgeIDs: knowledgeIDs,
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
func newAnswerabilityRetrieverWithHit() *fakeKnowledgeContextRetriever {
|
||||
return &fakeKnowledgeContextRetriever{
|
||||
knowledgeBaseIDs: []int64{1},
|
||||
result: &retrievers.KnowledgeRetrieveResult{
|
||||
KnowledgeBaseIDs: []int64{1},
|
||||
Hits: []rag.RetrieveResult{
|
||||
{KnowledgeBaseID: 1, DocumentID: 10, ChunkID: 101, Score: 0.93, Content: "购买后七天内且未使用可以退款。"},
|
||||
},
|
||||
ContextResults: []rag.RetrieveResult{
|
||||
{KnowledgeBaseID: 1, DocumentID: 10, ChunkID: 101, Score: 0.93, Content: "购买后七天内且未使用可以退款。"},
|
||||
},
|
||||
ContextText: "购买后七天内且未使用可以退款。",
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
type fakeKnowledgeContextRetriever struct {
|
||||
knowledgeBaseIDs []int64
|
||||
result *retrievers.KnowledgeRetrieveResult
|
||||
err error
|
||||
lastOptions retrievers.KnowledgeRetrieveOptions
|
||||
lastQuery string
|
||||
}
|
||||
|
||||
func (f *fakeKnowledgeContextRetriever) KnowledgeBaseIDs() []int64 {
|
||||
return append([]int64(nil), f.knowledgeBaseIDs...)
|
||||
}
|
||||
|
||||
func (f *fakeKnowledgeContextRetriever) RetrieveContextByOptions(ctx context.Context, opts retrievers.KnowledgeRetrieveOptions, query string) (*retrievers.KnowledgeRetrieveResult, error) {
|
||||
f.lastOptions = opts
|
||||
f.lastQuery = query
|
||||
return f.result, f.err
|
||||
}
|
||||
|
||||
type fakeAnswerabilityChatModel struct {
|
||||
response string
|
||||
err error
|
||||
input []*schema.Message
|
||||
}
|
||||
|
||||
func (f *fakeAnswerabilityChatModel) Generate(ctx context.Context, input []*schema.Message, opts ...model.Option) (*schema.Message, error) {
|
||||
f.input = input
|
||||
if f.err != nil {
|
||||
return nil, f.err
|
||||
}
|
||||
return schema.AssistantMessage(f.response, nil), nil
|
||||
}
|
||||
|
||||
func (f *fakeAnswerabilityChatModel) Stream(ctx context.Context, input []*schema.Message, opts ...model.Option) (*schema.StreamReader[*schema.Message], error) {
|
||||
return nil, errors.New("stream is not implemented in fakeAnswerabilityChatModel")
|
||||
}
|
||||
Reference in New Issue
Block a user