Files
ragflow/internal/ingestion/knowledge_compile/event.go
Zhichang Yu 01d667296d refactor(knowledge_compile): global compile pool, token-budget batching, and DocEngine-only deletion (#17679)
## Summary

This PR refactors the Go knowledge-compilation ingestion pipeline
(`internal/ingestion/knowledge_compile` +
`internal/ingestion/component/knowledge_compiler`) with three related
changes:

- **Token-budget batching for LLM merge decisions.**
`LLMMergeDecider.DecideBatch` previously stuffed every `(existing,
candidate)` pair into a single LLM call, risking `max_token` overflow.
It now splits pairs into token-bounded sub-batches (budget =
`llmMaxTokens * 0.85`) via `tokenizer.NumTokensFromString`, runs them
concurrently while preserving the global pair index, and never
reindexes.
- **Process-level global compile pool.** Introduces a single vCPU-sized
goroutine pool (`pool.go`, env `KC_COMPILE_CONCURRENCY`) dedicated to
*all* knowledge-compilation stages. KNN search loop, `DecideBatch`
sub-batches, `WriteMerged`/`DeleteMerged` internals, and the
component-level (structure/mindmap) per-call pools are all unified into
it via an injected submitter. No more per-job short-lived goroutines in
`runCompilerJobs` (futures are collected then awaited on the caller).
Fan-out stays bounded by the pool worker count; these stages are
docengine-bounded / LLM-bounded, not CPU-bounded.
- **DocEngine-only deletion.** `Consumer.processBatch` deletion no
longer loads the deleted docs' products into memory. Two sequential
DocEngine calls replace the old in-memory surgery:
- `DeleteDocLevelForDocs` — one `DeleteChunks` over `doc_id IN
deletedDocIDs` (merged rows carry `doc_id == kb`, so only per-doc
products match).
- `StripMergedSources` — one `Search` of `kc_merged=1` rows filtered by
`source_doc_ids IN deletedDocIDs` (intersection pushed down to the
engine), `UpdateChunks` the source array of survivors, and
`DeleteChunks` the rows whose array became empty.

## Changes

- `internal/ingestion/knowledge_compile/pool.go` (new): global
`compilerPool` +
`runCompilerJobs`/`SubmitCompilerJob`/`SubmitCompilerJobs`.
- `internal/ingestion/knowledge_compile/consumer.go`: deletion rewritten
to the two DocEngine calls;
`mergedBase`/`toDelete`/`stripDeletedSources` removed.
- `internal/ingestion/knowledge_compile/writer.go`:
`DeleteDocLevelForDocs` + `StripMergedSources` replace
`DeleteMergedForDoc`/`DeleteMerged`.
- `internal/ingestion/knowledge_compile/reader.go`: drop
`LoadMergedBySourceDoc` + `containsString` (keep `LoadDocProducts` for
the completion branch).
- `internal/ingestion/knowledge_compile/dedup.go`: `NewLLMDeduper` takes
`llmMaxTokens`; wires `SetMaxBatchTokens`/`SetSubmitter`.
- `internal/ingestion/knowledge_compiler/{structure,merge}.go`,
`mindmap/mindmap.go`, `pool_wiring.go`: token-budget split + submitter
injection.
- Tests: `structure_test.go` (token-budget split), `dedup_test.go`,
`consumer_test.go` (tombstone + DocEngine deletion assertions) updated.

## Validation

`bash build.sh --test -race ./internal/ingestion/knowledge_compile/...
./internal/ingestion/component/knowledge_compiler/...` passes (unit
tier, no external services).

🤖 Generated with [CodeBuddy](https://www.codebuddy.ai)

---------

Co-authored-by: yuzhichang <yuzhichang@infiniflow.ai>
2026-08-02 17:06:29 +08:00

125 lines
4.9 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 knowledge_compile implements the dataset-level post-processing consumer described in
// docs/develop/knowledge_compile_design.md §11 (Option E).
//
// Pipeline (KnowledgeCompiler, per document) writes compiled chunks with
// available_int=0. Its completion/deletion is recorded by appending a
// BacklogEntry to the KB's durable MySQL scheduling row
// (knowledge_compile_docs), then waking idle workers over NATS. A cluster of
// competing workers claims a closed batch (backlog -> inflight) per KB, runs
// dataset-level dedup on that batch, and writes the merged dataset-level
// products with available_int=1. The MySQL row — not the broker — is the
// scheduling system of record and the source of same-KB serialization.
package knowledge_compile
import (
"context"
"encoding/json"
)
// Subjects and event types for the knowledge-compile stream. Both sit under the
// knowledge.compile.events.> prefix declared on the NATS stream/consumer.
const (
SubjectCompleted = "knowledge.compile.events.completed"
SubjectDeleted = "knowledge.compile.events.deleted"
)
// EventType enumerates the KC event kinds.
type EventType string
const (
EventTypeCompleted EventType = "doc_completed"
EventTypeDeleted EventType = "doc_deleted"
)
// KCCompileEvent is the payload published when a document's pipeline finishes
// (doc_completed) or is deleted (doc_deleted). It is intentionally decoupled
// from common.TaskMessage because the consumer reads the raw JSON body.
type KCCompileEvent struct {
TenantID string `json:"tenant_id"`
DatasetID string `json:"dataset_id"` // the KB scope
DocID string `json:"doc_id"` // the contributing document
EventType string `json:"event_type"` // EventType value
Seq uint64 `json:"seq"` // per-doc monotonic sequence, for out-of-order correction
Timestamp int64 `json:"ts"`
}
// Subject returns the NATS subject for this event.
func (e KCCompileEvent) Subject() string {
if EventType(e.EventType) == EventTypeDeleted {
return SubjectDeleted
}
return SubjectCompleted
}
// Marshal serializes the event to JSON.
func (e KCCompileEvent) Marshal() ([]byte, error) { return json.Marshal(e) }
// ParseEvent deserializes a KCCompileEvent from raw message bytes.
func ParseEvent(data []byte) (KCCompileEvent, error) {
var e KCCompileEvent
if err := json.Unmarshal(data, &e); err != nil {
return KCCompileEvent{}, err
}
return e, nil
}
// defaultPublisher is the package-level Publisher used by the publishing path
// (PublishCompleted / PublishDeleted). It is installed by Provision (called
// once by the owning Ingestor at startup). Until then, publishing is a no-op.
var defaultPublisher Publisher
// defaultClaimer is the package-level Claimer handed to the consumer workers.
// It is the same underlying instance as defaultPublisher (a *mysqlScheduler
// satisfies both roles), set together by SetScheduler.
var defaultClaimer Claimer
// SetScheduler installs the package-level Publisher and Claimer used for
// publishing and consuming. The same *mysqlScheduler satisfies both interfaces.
func SetScheduler(s Publisher) {
defaultPublisher = s
if c, ok := s.(Claimer); ok {
defaultClaimer = c
}
}
// DefaultPublisher returns the package-level Publisher used by the producer path.
func DefaultPublisher() Publisher { return defaultPublisher }
// DefaultClaimer returns the package-level Claimer used by consumer workers.
func DefaultClaimer() Claimer { return defaultClaimer }
// PublishCompleted records a doc_completed event: it appends the doc to the
// KB's durable MySQL backlog and wakes idle workers over NATS. It is a no-op
// when no Publisher has been installed (e.g. DB unavailable). A failure is
// returned so callers can log but never fail the pipeline on it.
func PublishCompleted(ctx context.Context, tenantID, datasetID, docID string, seq uint64) error {
if defaultPublisher == nil {
return nil
}
return defaultPublisher.Publish(ctx, tenantID, datasetID, docID, string(EventTypeCompleted), seq)
}
// PublishDeleted records a doc_deleted event the same way (Publish handles the
// append + notify pairing).
func PublishDeleted(ctx context.Context, tenantID, datasetID, docID string, seq uint64) error {
if defaultPublisher == nil {
return nil
}
return defaultPublisher.Publish(ctx, tenantID, datasetID, docID, string(EventTypeDeleted), seq)
}