gitea/infio-copilot-dev
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update embeddings

Browse Source
This commit is contained in:
duanfuxiang
2025-06-14 09:17:44 +08:00
parent c71a13a659
commit f1ecc16c26
4 changed files with 322 additions and 211 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/components/inline-edit/InlineEdit.tsx
View File

@@ -258,6 +258,8 @@ export const InlineEdit: React.FC<InlineEditProps> = ({
let fileContent: string;
try {
fileContent = await plugin.app.vault.cachedRead(activeFile);
// 清理null字节，防止PostgreSQL UTF8编码错误
fileContent = fileContent.replace(/\0/g, '');
} catch (err) {
const error = err as Error;
console.error(t("inlineEdit.readFileError"), error.message);
@@ -278,7 +280,9 @@ export const InlineEdit: React.FC<InlineEditProps> = ({
return;
}
const oldContent = await plugin.app.vault.read(activeFile);
let oldContent = await plugin.app.vault.read(activeFile);
// 清理null字节，防止PostgreSQL UTF8编码错误
oldContent = oldContent.replace(/\0/g, '');
await plugin.app.workspace.getLeaf(true).setViewState({
type: APPLY_VIEW_TYPE,
active: true,

4
src/core/file-search/match/coreplugin-match.ts
View File

@@ -56,7 +56,9 @@ export async function matchSearchUsingCorePlugin(
break;
}
const content = await vault.cachedRead(file as TFile);
let content = await vault.cachedRead(file as TFile);
// 清理null字节，防止PostgreSQL UTF8编码错误
content = content.replace(/\0/g, '');
const lines = content.split('\n');
// `fileMatches.result.content` holds an array of matches for the file.

231
src/database/modules/vector/vector-manager.ts
View File

@@ -52,6 +52,29 @@ export class VectorManager {
)
}
// 强制垃圾回收的辅助方法
private forceGarbageCollection() {
try {
if (typeof global !== 'undefined' && global.gc) {
global.gc()
} else if (typeof window !== 'undefined' && (window as any).gc) {
(window as any).gc()
}
} catch (e) {
// 忽略垃圾回收错误
}
}
// 检查并清理内存的辅助方法
private async memoryCleanup(batchCount: number) {
// 每10批次强制垃圾回收
if (batchCount % 10 === 0) {
this.forceGarbageCollection()
// 短暂延迟让内存清理完成
await new Promise(resolve => setTimeout(resolve, 100))
}
}
async updateVaultIndex(
embeddingModel: EmbeddingModel,
options: {
@@ -100,10 +123,14 @@ export class VectorManager {
},
)
const skippedFiles: string[] = []
const contentChunks: InsertVector[] = (
await Promise.all(
filesToIndex.map(async (file) => {
const fileContent = await this.app.vault.cachedRead(file)
try {
let fileContent = await this.app.vault.cachedRead(file)
// 清理null字节，防止PostgreSQL UTF8编码错误
fileContent = fileContent.replace(/\0/g, '')
const fileDocuments = await textSplitter.createDocuments([
fileContent,
])
@@ -111,7 +138,7 @@ export class VectorManager {
return {
path: file.path,
mtime: file.stat.mtime,
content: chunk.pageContent,
content: chunk.pageContent.replace(/\0/g, ''), // 再次清理，确保安全
embedding: [],
metadata: {
startLine: Number(chunk.metadata.loc.lines.from),
@@ -119,10 +146,20 @@ export class VectorManager {
},
}
})
} catch (error) {
console.warn(`跳过文件 ${file.path}:`, error.message)
skippedFiles.push(file.path)
return []
}
}),
)
).flat()
if (skippedFiles.length > 0) {
console.warn(`跳过了 ${skippedFiles.length} 个有问题的文件:`, skippedFiles)
new Notice(`跳过了 ${skippedFiles.length} 个有问题的文件`)
}
updateProgress?.({
completedChunks: 0,
totalChunks: contentChunks.length,
@@ -130,18 +167,22 @@ export class VectorManager {
})
const embeddingProgress = { completed: 0 }
const embeddingChunks: InsertVector[] = []
const insertBatchSize = 64 // 数据库插入批量大小
// 减少批量大小以降低内存压力
const insertBatchSize = 16 // 从64降低到16
let batchCount = 0
try {
if (embeddingModel.supportsBatch) {
// 支持批量处理的提供商：使用批量处理逻辑
const embeddingBatchSize = 64 // API批量处理大小
// 支持批量处理的提供商：使用流式处理逻辑
const embeddingBatchSize = 16 // 从64降低到16
for (let i = 0; i < contentChunks.length; i += embeddingBatchSize) {
batchCount++
const batchChunks = contentChunks.slice(i, Math.min(i + embeddingBatchSize, contentChunks.length))
const batchTexts = batchChunks.map(chunk => chunk.content)
const embeddedBatch: InsertVector[] = []
await backOff(
async () => {
const batchEmbeddings = await embeddingModel.getBatchEmbeddings(batchTexts)
@@ -155,7 +196,22 @@ export class VectorManager {
embedding: batchEmbeddings[j],
metadata: batchChunks[j].metadata,
}
embeddingChunks.push(embeddedChunk)
embeddedBatch.push(embeddedChunk)
}
},
{
numOfAttempts: 3, // 减少重试次数
startingDelay: 500, // 减少延迟
timeMultiple: 1.5,
jitter: 'full',
},
)
// 立即插入当前批次，避免内存累积
if (embeddedBatch.length > 0) {
await this.repository.insertVectors(embeddedBatch, embeddingModel)
// 清理批次数据
embeddedBatch.length = 0
}
embeddingProgress.completed += batchChunks.length
@@ -164,20 +220,26 @@ export class VectorManager {
totalChunks: contentChunks.length,
totalFiles: filesToIndex.length,
})
},
{
numOfAttempts: 5,
startingDelay: 1000,
timeMultiple: 1.5,
jitter: 'full',
},
)
// 定期内存清理
await this.memoryCleanup(batchCount)
}
} else {
// 不支持批量处理的提供商：使用原来的逐个处理逻辑
const limit = pLimit(50)
// 不支持批量处理的提供商：使用流式处理逻辑
const limit = pLimit(10) // 从50降低到10，减少并发压力
const abortController = new AbortController()
const tasks = contentChunks.map((chunk) =>
// 流式处理：分批处理并立即插入
for (let i = 0; i < contentChunks.length; i += insertBatchSize) {
if (abortController.signal.aborted) {
throw new Error('Operation was aborted')
}
batchCount++
const batchChunks = contentChunks.slice(i, Math.min(i + insertBatchSize, contentChunks.length))
const embeddedBatch: InsertVector[] = []
const tasks = batchChunks.map((chunk) =>
limit(async () => {
if (abortController.signal.aborted) {
throw new Error('Operation was aborted')
@@ -193,17 +255,11 @@ export class VectorManager {
embedding,
metadata: chunk.metadata,
}
embeddingChunks.push(embeddedChunk)
embeddingProgress.completed++
updateProgress?.({
completedChunks: embeddingProgress.completed,
totalChunks: contentChunks.length,
totalFiles: filesToIndex.length,
})
embeddedBatch.push(embeddedChunk)
},
{
numOfAttempts: 5,
startingDelay: 1000,
numOfAttempts: 3, // 减少重试次数
startingDelay: 500, // 减少延迟
timeMultiple: 1.5,
jitter: 'full',
},
@@ -216,19 +272,23 @@ export class VectorManager {
)
await Promise.all(tasks)
// 立即插入当前批次
if (embeddedBatch.length > 0) {
await this.repository.insertVectors(embeddedBatch, embeddingModel)
// 清理批次数据
embeddedBatch.length = 0
}
// all embedding generated, batch insert
if (embeddingChunks.length > 0) {
// batch insert all vectors
let inserted = 0
while (inserted < embeddingChunks.length) {
const chunksToInsert = embeddingChunks.slice(
inserted,
Math.min(inserted + insertBatchSize, embeddingChunks.length)
)
await this.repository.insertVectors(chunksToInsert, embeddingModel)
inserted += chunksToInsert.length
embeddingProgress.completed += batchChunks.length
updateProgress?.({
completedChunks: embeddingProgress.completed,
totalChunks: contentChunks.length,
totalFiles: filesToIndex.length,
})
// 定期内存清理
await this.memoryCleanup(batchCount)
}
}
} catch (error) {
@@ -244,6 +304,9 @@ export class VectorManager {
console.error('Error embedding chunks:', error)
throw error
}
} finally {
// 最终清理
this.forceGarbageCollection()
}
}
@@ -252,7 +315,7 @@ export class VectorManager {
chunkSize: number,
file: TFile
) {
try {
// Delete existing vectors for the files
await this.repository.deleteVectorsForSingleFile(
file.path,
@@ -266,7 +329,9 @@ export class VectorManager {
chunkSize,
},
)
const fileContent = await this.app.vault.cachedRead(file)
let fileContent = await this.app.vault.cachedRead(file)
// 清理null字节，防止PostgreSQL UTF8编码错误
fileContent = fileContent.replace(/\0/g, '')
const fileDocuments = await textSplitter.createDocuments([
fileContent,
])
@@ -275,7 +340,7 @@ export class VectorManager {
return {
path: file.path,
mtime: file.stat.mtime,
content: chunk.pageContent,
content: chunk.pageContent.replace(/\0/g, ''), // 再次清理，确保安全
embedding: [],
metadata: {
startLine: Number(chunk.metadata.loc.lines.from),
@@ -284,19 +349,23 @@ export class VectorManager {
}
})
const embeddingChunks: InsertVector[] = []
const insertBatchSize = 64 // 数据库插入批量大小
// 减少批量大小以降低内存压力
const insertBatchSize = 16 // 从64降低到16
let batchCount = 0
try {
if (embeddingModel.supportsBatch) {
// 支持批量处理的提供商：使用批量处理逻辑
const embeddingBatchSize = 64 // API批量处理大小
// 支持批量处理的提供商：使用流式处理逻辑
const embeddingBatchSize = 16 // 从64降低到16
for (let i = 0; i < contentChunks.length; i += embeddingBatchSize) {
console.log(`Embedding batch ${i / embeddingBatchSize + 1} of ${Math.ceil(contentChunks.length / embeddingBatchSize)}`)
batchCount++
console.log(`Embedding batch ${batchCount} of ${Math.ceil(contentChunks.length / embeddingBatchSize)}`)
const batchChunks = contentChunks.slice(i, Math.min(i + embeddingBatchSize, contentChunks.length))
const batchTexts = batchChunks.map(chunk => chunk.content)
const embeddedBatch: InsertVector[] = []
await backOff(
async () => {
const batchEmbeddings = await embeddingModel.getBatchEmbeddings(batchTexts)
@@ -310,22 +379,43 @@ export class VectorManager {
embedding: batchEmbeddings[j],
metadata: batchChunks[j].metadata,
}
embeddingChunks.push(embeddedChunk)
embeddedBatch.push(embeddedChunk)
}
},
{
numOfAttempts: 5,
startingDelay: 1000,
numOfAttempts: 3, // 减少重试次数
startingDelay: 500, // 减少延迟
timeMultiple: 1.5,
jitter: 'full',
},
)
// 立即插入当前批次
if (embeddedBatch.length > 0) {
await this.repository.insertVectors(embeddedBatch, embeddingModel)
// 清理批次数据
embeddedBatch.length = 0
}
// 定期内存清理
await this.memoryCleanup(batchCount)
}
} else {
// 不支持批量处理的提供商：使用原来的逐个处理逻辑
const limit = pLimit(50)
// 不支持批量处理的提供商：使用流式处理逻辑
const limit = pLimit(10) // 从50降低到10
const abortController = new AbortController()
const tasks = contentChunks.map((chunk) =>
// 流式处理：分批处理并立即插入
for (let i = 0; i < contentChunks.length; i += insertBatchSize) {
if (abortController.signal.aborted) {
throw new Error('Operation was aborted')
}
batchCount++
const batchChunks = contentChunks.slice(i, Math.min(i + insertBatchSize, contentChunks.length))
const embeddedBatch: InsertVector[] = []
const tasks = batchChunks.map((chunk) =>
limit(async () => {
if (abortController.signal.aborted) {
throw new Error('Operation was aborted')
@@ -341,11 +431,11 @@ export class VectorManager {
embedding,
metadata: chunk.metadata,
}
embeddingChunks.push(embeddedChunk)
embeddedBatch.push(embeddedChunk)
},
{
numOfAttempts: 5,
startingDelay: 1000,
numOfAttempts: 3, // 减少重试次数
startingDelay: 500, // 减少延迟
timeMultiple: 1.5,
jitter: 'full',
},
@@ -358,19 +448,27 @@ export class VectorManager {
)
await Promise.all(tasks)
// 立即插入当前批次
if (embeddedBatch.length > 0) {
await this.repository.insertVectors(embeddedBatch, embeddingModel)
// 清理批次数据
embeddedBatch.length = 0
}
// all embedding generated, batch insert
if (embeddingChunks.length > 0) {
let inserted = 0
while (inserted < embeddingChunks.length) {
const chunksToInsert = embeddingChunks.slice(inserted, Math.min(inserted + insertBatchSize, embeddingChunks.length))
await this.repository.insertVectors(chunksToInsert, embeddingModel)
inserted += chunksToInsert.length
// 定期内存清理
await this.memoryCleanup(batchCount)
}
}
} catch (error) {
console.error('Error embedding chunks:', error)
} finally {
// 最终清理
this.forceGarbageCollection()
}
} catch (error) {
console.warn(`跳过文件 ${file.path}:`, error.message)
new Notice(`跳过文件 ${file.name}: ${error.message}`)
}
}
@@ -424,13 +522,16 @@ export class VectorManager {
// Check for updated or new files
filesToIndex = await Promise.all(
filesToIndex.map(async (file) => {
try {
const fileChunks = await this.repository.getVectorsByFilePath(
file.path,
embeddingModel,
)
if (fileChunks.length === 0) {
// File is not indexed, so we need to index it
const fileContent = await this.app.vault.cachedRead(file)
let fileContent = await this.app.vault.cachedRead(file)
// 清理null字节，防止PostgreSQL UTF8编码错误
fileContent = fileContent.replace(/\0/g, '')
if (fileContent.length === 0) {
// Ignore empty files
return null
@@ -443,6 +544,10 @@ export class VectorManager {
return file
}
return null
} catch (error) {
console.warn(`跳过文件 ${file.path}:`, error.message)
return null
}
}),
).then((files) => files.filter(Boolean))

2
src/database/modules/vector/vector-repository.ts
View File

@@ -102,7 +102,7 @@ export class VectorRepository {
const params = data.flatMap(vector => [
vector.path,
vector.mtime,
vector.content,
vector.content.replace(/\0/g, ''), // 清理null字节
`[${vector.embedding.join(',')}]`, // 转换为PostgreSQL vector格式
vector.metadata
])