@deepagents/retrieval 7.2.0 → 7.4.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/index.d.ts +7 -7
- package/dist/index.js +561 -7
- package/dist/index.js.map +4 -4
- package/dist/lib/connectors/index.d.ts +7 -7
- package/dist/lib/connectors/index.js +649 -7
- package/dist/lib/connectors/index.js.map +4 -4
- package/dist/lib/connectors/local.d.ts +1 -1
- package/dist/lib/connectors/local.d.ts.map +1 -1
- package/dist/lib/connectors/pdf.d.ts +1 -1
- package/dist/lib/connectors/pdf.d.ts.map +1 -1
- package/dist/lib/connectors/repo.d.ts +1 -1
- package/dist/lib/embedders/fastembed.d.ts +1 -1
- package/dist/lib/embedders/huggingface.d.ts +1 -1
- package/dist/lib/ingest.d.ts +3 -3
- package/dist/lib/similiarty-search.d.ts +1 -1
- package/dist/lib/stores/sqlite/ddl.sqlite.d.ts +1 -1
- package/dist/lib/stores/sqlite/ddl.sqlite.d.ts.map +1 -1
- package/dist/lib/stores/sqlite/node-sqlite.d.ts +1 -1
- package/dist/lib/stores/sqlite/open.d.ts +1 -2
- package/dist/lib/stores/sqlite/open.d.ts.map +1 -1
- package/dist/lib/stores/sqlite/sqlite.d.ts +1 -1
- package/package.json +4 -3
- package/dist/lib/connectors/connector.js +0 -1
- package/dist/lib/connectors/connector.js.map +0 -7
- package/dist/lib/connectors/github.js +0 -164
- package/dist/lib/connectors/github.js.map +0 -7
- package/dist/lib/connectors/linear.js +0 -29
- package/dist/lib/connectors/linear.js.map +0 -7
- package/dist/lib/connectors/local.js +0 -98
- package/dist/lib/connectors/local.js.map +0 -7
- package/dist/lib/connectors/pdf.js +0 -58
- package/dist/lib/connectors/pdf.js.map +0 -7
- package/dist/lib/connectors/repo.js +0 -179
- package/dist/lib/connectors/repo.js.map +0 -7
- package/dist/lib/connectors/rss.js +0 -128
- package/dist/lib/connectors/rss.js.map +0 -7
- package/dist/lib/embedders/fastembed.js +0 -63
- package/dist/lib/embedders/fastembed.js.map +0 -7
- package/dist/lib/embedders/huggingface.js +0 -42
- package/dist/lib/embedders/huggingface.js.map +0 -7
- package/dist/lib/ingest.js +0 -117
- package/dist/lib/ingest.js.map +0 -7
- package/dist/lib/pipeline.js +0 -1
- package/dist/lib/pipeline.js.map +0 -7
- package/dist/lib/sidecar.js +0 -1
- package/dist/lib/sidecar.js.map +0 -7
- package/dist/lib/similiarty-search.js +0 -42
- package/dist/lib/similiarty-search.js.map +0 -7
- package/dist/lib/stores/cid.js +0 -8
- package/dist/lib/stores/cid.js.map +0 -7
- package/dist/lib/stores/sqlite/ddl.sqlite.js +0 -53
- package/dist/lib/stores/sqlite/ddl.sqlite.js.map +0 -7
- package/dist/lib/stores/sqlite/node-sqlite.js +0 -9
- package/dist/lib/stores/sqlite/node-sqlite.js.map +0 -7
- package/dist/lib/stores/sqlite/open.js +0 -11
- package/dist/lib/stores/sqlite/open.js.map +0 -7
- package/dist/lib/stores/sqlite/sqlite.js +0 -245
- package/dist/lib/stores/sqlite/sqlite.js.map +0 -7
- package/dist/lib/stores/store.js +0 -1
- package/dist/lib/stores/store.js.map +0 -7
package/dist/lib/ingest.js
DELETED
|
@@ -1,117 +0,0 @@
|
|
|
1
|
-
import {
|
|
2
|
-
MarkdownTextSplitter,
|
|
3
|
-
RecursiveCharacterTextSplitter
|
|
4
|
-
} from "@langchain/textsplitters";
|
|
5
|
-
import { cid } from "./stores/cid.js";
|
|
6
|
-
async function ingest(config, callback) {
|
|
7
|
-
const splitter = config.splitter ?? split;
|
|
8
|
-
const embedder = config.embedder;
|
|
9
|
-
const corpuses = config.connector.sources();
|
|
10
|
-
const { expiresAfter } = config.connector;
|
|
11
|
-
const expiryDate = expiresAfter === void 0 ? void 0 : new Date(Date.now() + expiresAfter);
|
|
12
|
-
for await (const it of corpuses) {
|
|
13
|
-
callback?.(it.id);
|
|
14
|
-
const content = await it.content();
|
|
15
|
-
if (!content.trim()) {
|
|
16
|
-
continue;
|
|
17
|
-
}
|
|
18
|
-
await config.store.index(
|
|
19
|
-
config.connector.sourceId,
|
|
20
|
-
{
|
|
21
|
-
id: it.id,
|
|
22
|
-
cid: cid(content),
|
|
23
|
-
metadata: it.metadata,
|
|
24
|
-
chunker: async function* () {
|
|
25
|
-
const values = await splitter(it.id, content);
|
|
26
|
-
const batchSize = 40;
|
|
27
|
-
for (let i = 0; i < values.length; i += batchSize) {
|
|
28
|
-
const batch = values.slice(i, i + batchSize);
|
|
29
|
-
const { embeddings } = await embedder(batch);
|
|
30
|
-
for (let j = 0; j < embeddings.length; j++) {
|
|
31
|
-
yield {
|
|
32
|
-
content: batch[j],
|
|
33
|
-
embedding: embeddings[j]
|
|
34
|
-
};
|
|
35
|
-
}
|
|
36
|
-
}
|
|
37
|
-
}
|
|
38
|
-
},
|
|
39
|
-
expiryDate
|
|
40
|
-
);
|
|
41
|
-
}
|
|
42
|
-
}
|
|
43
|
-
function split(id, content) {
|
|
44
|
-
const splitter = new MarkdownTextSplitter();
|
|
45
|
-
return splitter.splitText(content);
|
|
46
|
-
}
|
|
47
|
-
function normalizeNewlines(value) {
|
|
48
|
-
return value.replace(/\r\n/g, "\n");
|
|
49
|
-
}
|
|
50
|
-
function computePositions(originalContent, chunks) {
|
|
51
|
-
if (!chunks.length) {
|
|
52
|
-
return [];
|
|
53
|
-
}
|
|
54
|
-
const normalizedContent = normalizeNewlines(originalContent);
|
|
55
|
-
const positions = [];
|
|
56
|
-
let searchOffset = 0;
|
|
57
|
-
for (const chunk of chunks) {
|
|
58
|
-
const normalizedChunk = normalizeNewlines(chunk);
|
|
59
|
-
const trimmedChunk = normalizedChunk.trim();
|
|
60
|
-
const seek = (needle, fromIndex) => needle ? normalizedContent.indexOf(needle, fromIndex) : -1;
|
|
61
|
-
let matchIndex = seek(normalizedChunk, searchOffset);
|
|
62
|
-
let matchValue = normalizedChunk;
|
|
63
|
-
if (matchIndex === -1 && trimmedChunk) {
|
|
64
|
-
matchIndex = seek(trimmedChunk, searchOffset);
|
|
65
|
-
matchValue = trimmedChunk;
|
|
66
|
-
}
|
|
67
|
-
if (matchIndex === -1) {
|
|
68
|
-
matchIndex = seek(normalizedChunk, 0);
|
|
69
|
-
matchValue = normalizedChunk;
|
|
70
|
-
}
|
|
71
|
-
if (matchIndex === -1 && trimmedChunk) {
|
|
72
|
-
matchIndex = seek(trimmedChunk, 0);
|
|
73
|
-
matchValue = trimmedChunk;
|
|
74
|
-
}
|
|
75
|
-
if (matchIndex === -1) {
|
|
76
|
-
positions.push(null);
|
|
77
|
-
continue;
|
|
78
|
-
}
|
|
79
|
-
const before = normalizedContent.slice(0, matchIndex);
|
|
80
|
-
const beforeLines = before.split("\n");
|
|
81
|
-
const startLine = beforeLines.length;
|
|
82
|
-
const startColumn = beforeLines[beforeLines.length - 1].length + 1;
|
|
83
|
-
const lines = matchValue.split("\n");
|
|
84
|
-
const endLine = startLine + lines.length - 1;
|
|
85
|
-
const endColumn = lines.length === 1 ? startColumn + lines[0].length : lines[lines.length - 1].length + 1;
|
|
86
|
-
positions.push({ startLine, startColumn, endLine, endColumn });
|
|
87
|
-
searchOffset = matchIndex + matchValue.length;
|
|
88
|
-
}
|
|
89
|
-
return positions;
|
|
90
|
-
}
|
|
91
|
-
function buildChunksWithPositions(originalContent, chunks) {
|
|
92
|
-
const positions = computePositions(originalContent, chunks);
|
|
93
|
-
return chunks.map((content, index) => ({
|
|
94
|
-
content,
|
|
95
|
-
index,
|
|
96
|
-
position: positions[index] ?? null
|
|
97
|
-
}));
|
|
98
|
-
}
|
|
99
|
-
async function splitTypeScriptWithPositions(id, content) {
|
|
100
|
-
const splitter = RecursiveCharacterTextSplitter.fromLanguage("js", {
|
|
101
|
-
chunkSize: 512,
|
|
102
|
-
chunkOverlap: 100
|
|
103
|
-
});
|
|
104
|
-
const docs = await splitter.createDocuments([content]);
|
|
105
|
-
const chunks = docs.map((d) => d.pageContent);
|
|
106
|
-
return buildChunksWithPositions(content, chunks);
|
|
107
|
-
}
|
|
108
|
-
async function splitTypeScript(id, content) {
|
|
109
|
-
const chunks = await splitTypeScriptWithPositions(id, content);
|
|
110
|
-
return chunks.map((chunk) => chunk.content);
|
|
111
|
-
}
|
|
112
|
-
export {
|
|
113
|
-
ingest,
|
|
114
|
-
splitTypeScript,
|
|
115
|
-
splitTypeScriptWithPositions
|
|
116
|
-
};
|
|
117
|
-
//# sourceMappingURL=ingest.js.map
|
package/dist/lib/ingest.js.map
DELETED
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../src/lib/ingest.ts"],
|
|
4
|
-
"sourcesContent": ["import {\n MarkdownTextSplitter,\n RecursiveCharacterTextSplitter,\n} from '@langchain/textsplitters';\n\nimport type { Connector } from './connectors/connector.js';\nimport type { Splitter } from './pipeline.js';\nimport { cid } from './stores/cid.js';\nimport type { Embedder, Store } from './stores/store.js';\n\nexport interface IngestionConfig {\n connector: Connector;\n store: Store;\n splitter?: Splitter;\n embedder: Embedder;\n}\n\nexport async function ingest(\n config: IngestionConfig,\n callback?: (it: string) => void,\n) {\n const splitter = config.splitter ?? split;\n const embedder = config.embedder;\n const corpuses = config.connector.sources();\n const { expiresAfter } = config.connector;\n const expiryDate =\n expiresAfter === undefined\n ? undefined\n : new Date(Date.now() + expiresAfter);\n\n for await (const it of corpuses) {\n callback?.(it.id);\n const content = await it.content();\n if (!content.trim()) {\n // skip empty files\n continue;\n }\n await config.store.index(\n config.connector.sourceId,\n {\n id: it.id,\n cid: cid(content),\n metadata: it.metadata,\n chunker: async function* () {\n // Embed in small batches to control memory usage\n const values = await splitter(it.id, content);\n const batchSize = 40;\n for (let i = 0; i < values.length; i += batchSize) {\n const batch = values.slice(i, i + batchSize);\n const { embeddings } = await embedder(batch);\n for (let j = 0; j < embeddings.length; j++) {\n yield {\n content: batch[j],\n embedding: embeddings[j],\n };\n }\n }\n },\n },\n expiryDate,\n );\n }\n}\n\nfunction split(id: string, content: string) {\n const splitter = new MarkdownTextSplitter();\n return splitter.splitText(content);\n}\n\nexport type ChunkPosition = {\n startLine: number;\n startColumn: number;\n endLine: number;\n endColumn: number;\n};\n\nexport type SplitChunkWithPosition = {\n content: string;\n position: ChunkPosition | null;\n index: number;\n};\n\nfunction normalizeNewlines(value: string) {\n return value.replace(/\\r\\n/g, '\\n');\n}\n\nfunction computePositions(\n originalContent: string,\n chunks: string[],\n): Array<ChunkPosition | null> {\n if (!chunks.length) {\n return [];\n }\n\n const normalizedContent = normalizeNewlines(originalContent);\n const positions: Array<ChunkPosition | null> = [];\n let searchOffset = 0;\n\n for (const chunk of chunks) {\n const normalizedChunk = normalizeNewlines(chunk);\n const trimmedChunk = normalizedChunk.trim();\n\n const seek = (needle: string, fromIndex: number) =>\n needle ? normalizedContent.indexOf(needle, fromIndex) : -1;\n\n let matchIndex = seek(normalizedChunk, searchOffset);\n let matchValue = normalizedChunk;\n\n if (matchIndex === -1 && trimmedChunk) {\n matchIndex = seek(trimmedChunk, searchOffset);\n matchValue = trimmedChunk;\n }\n\n if (matchIndex === -1) {\n matchIndex = seek(normalizedChunk, 0);\n matchValue = normalizedChunk;\n }\n\n if (matchIndex === -1 && trimmedChunk) {\n matchIndex = seek(trimmedChunk, 0);\n matchValue = trimmedChunk;\n }\n\n if (matchIndex === -1) {\n positions.push(null);\n continue;\n }\n\n const before = normalizedContent.slice(0, matchIndex);\n const beforeLines = before.split('\\n');\n const startLine = beforeLines.length;\n const startColumn = beforeLines[beforeLines.length - 1].length + 1;\n\n const lines = matchValue.split('\\n');\n const endLine = startLine + lines.length - 1;\n const endColumn =\n lines.length === 1\n ? startColumn + lines[0].length\n : lines[lines.length - 1].length + 1;\n\n positions.push({ startLine, startColumn, endLine, endColumn });\n searchOffset = matchIndex + matchValue.length;\n }\n\n return positions;\n}\n\nfunction buildChunksWithPositions(\n originalContent: string,\n chunks: string[],\n): SplitChunkWithPosition[] {\n const positions = computePositions(originalContent, chunks);\n return chunks.map((content, index) => ({\n content,\n index,\n position: positions[index] ?? null,\n }));\n}\n\nexport async function splitTypeScriptWithPositions(\n id: string,\n content: string,\n): Promise<SplitChunkWithPosition[]> {\n const splitter = RecursiveCharacterTextSplitter.fromLanguage('js', {\n chunkSize: 512,\n chunkOverlap: 100,\n });\n const docs = await splitter.createDocuments([content]);\n const chunks = docs.map((d) => d.pageContent);\n return buildChunksWithPositions(content, chunks);\n}\n\nexport async function splitTypeScript(id: string, content: string) {\n const chunks = await splitTypeScriptWithPositions(id, content);\n return chunks.map((chunk) => chunk.content);\n}\n"],
|
|
5
|
-
"mappings": "AAAA;AAAA,EACE;AAAA,EACA;AAAA,OACK;AAIP,SAAS,WAAW;AAUpB,eAAsB,OACpB,QACA,UACA;AACA,QAAM,WAAW,OAAO,YAAY;AACpC,QAAM,WAAW,OAAO;AACxB,QAAM,WAAW,OAAO,UAAU,QAAQ;AAC1C,QAAM,EAAE,aAAa,IAAI,OAAO;AAChC,QAAM,aACJ,iBAAiB,SACb,SACA,IAAI,KAAK,KAAK,IAAI,IAAI,YAAY;AAExC,mBAAiB,MAAM,UAAU;AAC/B,eAAW,GAAG,EAAE;AAChB,UAAM,UAAU,MAAM,GAAG,QAAQ;AACjC,QAAI,CAAC,QAAQ,KAAK,GAAG;AAEnB;AAAA,IACF;AACA,UAAM,OAAO,MAAM;AAAA,MACjB,OAAO,UAAU;AAAA,MACjB;AAAA,QACE,IAAI,GAAG;AAAA,QACP,KAAK,IAAI,OAAO;AAAA,QAChB,UAAU,GAAG;AAAA,QACb,SAAS,mBAAmB;AAE1B,gBAAM,SAAS,MAAM,SAAS,GAAG,IAAI,OAAO;AAC5C,gBAAM,YAAY;AAClB,mBAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK,WAAW;AACjD,kBAAM,QAAQ,OAAO,MAAM,GAAG,IAAI,SAAS;AAC3C,kBAAM,EAAE,WAAW,IAAI,MAAM,SAAS,KAAK;AAC3C,qBAAS,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC1C,oBAAM;AAAA,gBACJ,SAAS,MAAM,CAAC;AAAA,gBAChB,WAAW,WAAW,CAAC;AAAA,cACzB;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,MAAM,IAAY,SAAiB;AAC1C,QAAM,WAAW,IAAI,qBAAqB;AAC1C,SAAO,SAAS,UAAU,OAAO;AACnC;AAeA,SAAS,kBAAkB,OAAe;AACxC,SAAO,MAAM,QAAQ,SAAS,IAAI;AACpC;AAEA,SAAS,iBACP,iBACA,QAC6B;AAC7B,MAAI,CAAC,OAAO,QAAQ;AAClB,WAAO,CAAC;AAAA,EACV;AAEA,QAAM,oBAAoB,kBAAkB,eAAe;AAC3D,QAAM,YAAyC,CAAC;AAChD,MAAI,eAAe;AAEnB,aAAW,SAAS,QAAQ;AAC1B,UAAM,kBAAkB,kBAAkB,KAAK;AAC/C,UAAM,eAAe,gBAAgB,KAAK;AAE1C,UAAM,OAAO,CAAC,QAAgB,cAC5B,SAAS,kBAAkB,QAAQ,QAAQ,SAAS,IAAI;AAE1D,QAAI,aAAa,KAAK,iBAAiB,YAAY;AACnD,QAAI,aAAa;AAEjB,QAAI,eAAe,MAAM,cAAc;AACrC,mBAAa,KAAK,cAAc,YAAY;AAC5C,mBAAa;AAAA,IACf;AAEA,QAAI,eAAe,IAAI;AACrB,mBAAa,KAAK,iBAAiB,CAAC;AACpC,mBAAa;AAAA,IACf;AAEA,QAAI,eAAe,MAAM,cAAc;AACrC,mBAAa,KAAK,cAAc,CAAC;AACjC,mBAAa;AAAA,IACf;AAEA,QAAI,eAAe,IAAI;AACrB,gBAAU,KAAK,IAAI;AACnB;AAAA,IACF;AAEA,UAAM,SAAS,kBAAkB,MAAM,GAAG,UAAU;AACpD,UAAM,cAAc,OAAO,MAAM,IAAI;AACrC,UAAM,YAAY,YAAY;AAC9B,UAAM,cAAc,YAAY,YAAY,SAAS,CAAC,EAAE,SAAS;AAEjE,UAAM,QAAQ,WAAW,MAAM,IAAI;AACnC,UAAM,UAAU,YAAY,MAAM,SAAS;AAC3C,UAAM,YACJ,MAAM,WAAW,IACb,cAAc,MAAM,CAAC,EAAE,SACvB,MAAM,MAAM,SAAS,CAAC,EAAE,SAAS;AAEvC,cAAU,KAAK,EAAE,WAAW,aAAa,SAAS,UAAU,CAAC;AAC7D,mBAAe,aAAa,WAAW;AAAA,EACzC;AAEA,SAAO;AACT;AAEA,SAAS,yBACP,iBACA,QAC0B;AAC1B,QAAM,YAAY,iBAAiB,iBAAiB,MAAM;AAC1D,SAAO,OAAO,IAAI,CAAC,SAAS,WAAW;AAAA,IACrC;AAAA,IACA;AAAA,IACA,UAAU,UAAU,KAAK,KAAK;AAAA,EAChC,EAAE;AACJ;AAEA,eAAsB,6BACpB,IACA,SACmC;AACnC,QAAM,WAAW,+BAA+B,aAAa,MAAM;AAAA,IACjE,WAAW;AAAA,IACX,cAAc;AAAA,EAChB,CAAC;AACD,QAAM,OAAO,MAAM,SAAS,gBAAgB,CAAC,OAAO,CAAC;AACrD,QAAM,SAAS,KAAK,IAAI,CAAC,MAAM,EAAE,WAAW;AAC5C,SAAO,yBAAyB,SAAS,MAAM;AACjD;AAEA,eAAsB,gBAAgB,IAAY,SAAiB;AACjE,QAAM,SAAS,MAAM,6BAA6B,IAAI,OAAO;AAC7D,SAAO,OAAO,IAAI,CAAC,UAAU,MAAM,OAAO;AAC5C;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
package/dist/lib/pipeline.js
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
//# sourceMappingURL=pipeline.js.map
|
package/dist/lib/pipeline.js.map
DELETED
package/dist/lib/sidecar.js
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
//# sourceMappingURL=sidecar.js.map
|
package/dist/lib/sidecar.js.map
DELETED
|
@@ -1,42 +0,0 @@
|
|
|
1
|
-
import { embedMany } from "ai";
|
|
2
|
-
import { ingest } from "./ingest.js";
|
|
3
|
-
async function similaritySearch(query, config) {
|
|
4
|
-
const mode = config.connector.ingestWhen ?? "contentChanged";
|
|
5
|
-
let shouldIngest = true;
|
|
6
|
-
if (mode === "never") {
|
|
7
|
-
if (await config.store.sourceExists(config.connector.sourceId)) {
|
|
8
|
-
console.log(
|
|
9
|
-
`Skipping ingestion for source ${config.connector.sourceId} (ingestWhen=never and source exists)`
|
|
10
|
-
);
|
|
11
|
-
shouldIngest = false;
|
|
12
|
-
}
|
|
13
|
-
} else if (mode === "expired") {
|
|
14
|
-
const sourceExists = await config.store.sourceExists(
|
|
15
|
-
config.connector.sourceId
|
|
16
|
-
);
|
|
17
|
-
if (sourceExists && !await config.store.sourceExpired(config.connector.sourceId)) {
|
|
18
|
-
console.log(
|
|
19
|
-
`Skipping ingestion for source ${config.connector.sourceId} (ingestWhen=expired and source not expired)`
|
|
20
|
-
);
|
|
21
|
-
shouldIngest = false;
|
|
22
|
-
}
|
|
23
|
-
}
|
|
24
|
-
if (shouldIngest) {
|
|
25
|
-
await ingest(config);
|
|
26
|
-
}
|
|
27
|
-
return config.store.search(
|
|
28
|
-
query,
|
|
29
|
-
{ sourceId: config.connector.sourceId, topN: 50 },
|
|
30
|
-
config.embedder
|
|
31
|
-
).then(
|
|
32
|
-
(results) => results.map((it) => ({
|
|
33
|
-
...it,
|
|
34
|
-
similarity: 1 - it.distance,
|
|
35
|
-
distance: it.distance
|
|
36
|
-
}))
|
|
37
|
-
);
|
|
38
|
-
}
|
|
39
|
-
export {
|
|
40
|
-
similaritySearch
|
|
41
|
-
};
|
|
42
|
-
//# sourceMappingURL=similiarty-search.js.map
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../src/lib/similiarty-search.ts"],
|
|
4
|
-
"sourcesContent": ["import { embedMany } from 'ai';\n\nimport { type IngestionConfig, ingest } from './ingest.js';\n\nexport async function similaritySearch(\n query: string,\n config: Omit<IngestionConfig, 'splitter'>,\n) {\n // Ingest if needed, then perform vector similarity search via the configured store\n const mode = config.connector.ingestWhen ?? 'contentChanged';\n let shouldIngest = true;\n\n if (mode === 'never') {\n if (await config.store.sourceExists(config.connector.sourceId)) {\n console.log(\n `Skipping ingestion for source ${config.connector.sourceId} (ingestWhen=never and source exists)`,\n );\n shouldIngest = false;\n }\n } else if (mode === 'expired') {\n const sourceExists = await config.store.sourceExists(\n config.connector.sourceId,\n );\n if (\n sourceExists &&\n !(await config.store.sourceExpired(config.connector.sourceId))\n ) {\n console.log(\n `Skipping ingestion for source ${config.connector.sourceId} (ingestWhen=expired and source not expired)`,\n );\n shouldIngest = false;\n }\n }\n\n if (shouldIngest) {\n await ingest(config);\n }\n return config.store\n .search(\n query,\n { sourceId: config.connector.sourceId, topN: 50 },\n config.embedder,\n )\n .then((results) =>\n results.map((it) => ({\n ...it,\n similarity: 1 - it.distance,\n distance: it.distance,\n })),\n );\n}\n"],
|
|
5
|
-
"mappings": "AAAA,SAAS,iBAAiB;AAE1B,SAA+B,cAAc;AAE7C,eAAsB,iBACpB,OACA,QACA;AAEA,QAAM,OAAO,OAAO,UAAU,cAAc;AAC5C,MAAI,eAAe;AAEnB,MAAI,SAAS,SAAS;AACpB,QAAI,MAAM,OAAO,MAAM,aAAa,OAAO,UAAU,QAAQ,GAAG;AAC9D,cAAQ;AAAA,QACN,iCAAiC,OAAO,UAAU,QAAQ;AAAA,MAC5D;AACA,qBAAe;AAAA,IACjB;AAAA,EACF,WAAW,SAAS,WAAW;AAC7B,UAAM,eAAe,MAAM,OAAO,MAAM;AAAA,MACtC,OAAO,UAAU;AAAA,IACnB;AACA,QACE,gBACA,CAAE,MAAM,OAAO,MAAM,cAAc,OAAO,UAAU,QAAQ,GAC5D;AACA,cAAQ;AAAA,QACN,iCAAiC,OAAO,UAAU,QAAQ;AAAA,MAC5D;AACA,qBAAe;AAAA,IACjB;AAAA,EACF;AAEA,MAAI,cAAc;AAChB,UAAM,OAAO,MAAM;AAAA,EACrB;AACA,SAAO,OAAO,MACX;AAAA,IACC;AAAA,IACA,EAAE,UAAU,OAAO,UAAU,UAAU,MAAM,GAAG;AAAA,IAChD,OAAO;AAAA,EACT,EACC;AAAA,IAAK,CAAC,YACL,QAAQ,IAAI,CAAC,QAAQ;AAAA,MACnB,GAAG;AAAA,MACH,YAAY,IAAI,GAAG;AAAA,MACnB,UAAU,GAAG;AAAA,IACf,EAAE;AAAA,EACJ;AACJ;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
package/dist/lib/stores/cid.js
DELETED
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../../src/lib/stores/cid.ts"],
|
|
4
|
-
"sourcesContent": ["import { createHash } from 'node:crypto';\n\nexport function cid(content: string) {\n return createHash('sha256').update(content).digest('hex');\n}\n"],
|
|
5
|
-
"mappings": "AAAA,SAAS,kBAAkB;AAEpB,SAAS,IAAI,SAAiB;AACnC,SAAO,WAAW,QAAQ,EAAE,OAAO,OAAO,EAAE,OAAO,KAAK;AAC1D;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
|
@@ -1,53 +0,0 @@
|
|
|
1
|
-
const SETUP = `
|
|
2
|
-
PRAGMA busy_timeout = 5000;
|
|
3
|
-
PRAGMA page_size = 32768;
|
|
4
|
-
PRAGMA journal_mode = WAL;
|
|
5
|
-
PRAGMA synchronous = NORMAL;
|
|
6
|
-
PRAGMA temp_store = MEMORY;
|
|
7
|
-
PRAGMA cache_size = -131072; -- ~128 MiB page cache for faster repeated reads
|
|
8
|
-
PRAGMA mmap_size = 268435456; -- 256 MiB memory map window to cut syscalls
|
|
9
|
-
PRAGMA wal_autocheckpoint = 1000;
|
|
10
|
-
PRAGMA optimize;
|
|
11
|
-
`;
|
|
12
|
-
function vectors(dimension) {
|
|
13
|
-
return `CREATE VIRTUAL TABLE vec_chunks USING vec0(
|
|
14
|
-
source_id TEXT PARTITION KEY,
|
|
15
|
-
document_id TEXT,
|
|
16
|
-
embedding FLOAT[${dimension}] DISTANCE_METRIC=cosine,
|
|
17
|
-
+content TEXT -- auxiliary payload (not filterable) for reconstruction
|
|
18
|
-
)`;
|
|
19
|
-
}
|
|
20
|
-
const statements = [
|
|
21
|
-
`CREATE TABLE sources (
|
|
22
|
-
source_id TEXT PRIMARY KEY,
|
|
23
|
-
updated_at TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ','now')),
|
|
24
|
-
expires_at TEXT
|
|
25
|
-
) STRICT`,
|
|
26
|
-
`CREATE TABLE documents (
|
|
27
|
-
id TEXT PRIMARY KEY,
|
|
28
|
-
source_id TEXT NOT NULL,
|
|
29
|
-
cid TEXT NOT NULL,
|
|
30
|
-
metadata TEXT, -- JSON blob with arbitrary document metadata
|
|
31
|
-
updated_at TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ','now')),
|
|
32
|
-
FOREIGN KEY (source_id) REFERENCES sources(source_id) ON DELETE CASCADE
|
|
33
|
-
) STRICT`,
|
|
34
|
-
// Indexes to accelerate lookups / maintenance
|
|
35
|
-
`CREATE INDEX idx_documents_source ON documents(source_id)`,
|
|
36
|
-
`CREATE INDEX idx_documents_source_updated ON documents(source_id, updated_at)`,
|
|
37
|
-
`CREATE INDEX idx_documents_cid ON documents(cid)`,
|
|
38
|
-
`CREATE INDEX idx_sources_expires_at ON sources(expires_at) WHERE expires_at IS NOT NULL`,
|
|
39
|
-
// emulate cascade for the virtual table
|
|
40
|
-
`CREATE TRIGGER trg_documents_delete_vec
|
|
41
|
-
AFTER DELETE ON documents
|
|
42
|
-
BEGIN
|
|
43
|
-
DELETE FROM vec_chunks
|
|
44
|
-
WHERE source_id = OLD.source_id
|
|
45
|
-
AND document_id = OLD.id;
|
|
46
|
-
END`
|
|
47
|
-
];
|
|
48
|
-
export {
|
|
49
|
-
SETUP,
|
|
50
|
-
statements,
|
|
51
|
-
vectors
|
|
52
|
-
};
|
|
53
|
-
//# sourceMappingURL=ddl.sqlite.js.map
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../../../src/lib/stores/sqlite/ddl.sqlite.ts"],
|
|
4
|
-
"sourcesContent": ["// Embedding store schema for SQLite. SQLite stores the text of each CREATE and\n// @zukhruf/sqlite-drift compares it, so the text after each object name must\n// stay byte-identical to what earlier releases wrote, inline comments included.\n\n// Session-level PRAGMAs, run on each connection before the schema is applied.\nexport const SETUP = `\nPRAGMA busy_timeout = 5000;\nPRAGMA page_size = 32768;\nPRAGMA journal_mode = WAL;\nPRAGMA synchronous = NORMAL;\nPRAGMA temp_store = MEMORY;\nPRAGMA cache_size = -131072; -- ~128 MiB page cache for faster repeated reads\nPRAGMA mmap_size = 268435456; -- 256 MiB memory map window to cut syscalls\nPRAGMA wal_autocheckpoint = 1000;\nPRAGMA optimize;\n`;\n\nexport function vectors(dimension: number) {\n return `CREATE VIRTUAL TABLE vec_chunks USING vec0(\n source_id TEXT PARTITION KEY,\n document_id TEXT,\n embedding FLOAT[${dimension}] DISTANCE_METRIC=cosine,\n +content TEXT -- auxiliary payload (not filterable) for reconstruction\n)`;\n}\n\nexport const statements = [\n `CREATE TABLE sources (\n source_id TEXT PRIMARY KEY,\n updated_at TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ','now')),\n expires_at TEXT\n) STRICT`,\n `CREATE TABLE documents (\n id TEXT PRIMARY KEY,\n source_id TEXT NOT NULL,\n cid TEXT NOT NULL,\n metadata TEXT, -- JSON blob with arbitrary document metadata\n updated_at TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ','now')),\n FOREIGN KEY (source_id) REFERENCES sources(source_id) ON DELETE CASCADE\n) STRICT`,\n // Indexes to accelerate lookups / maintenance\n `CREATE INDEX idx_documents_source ON documents(source_id)`,\n `CREATE INDEX idx_documents_source_updated ON documents(source_id, updated_at)`,\n `CREATE INDEX idx_documents_cid ON documents(cid)`,\n `CREATE INDEX idx_sources_expires_at ON sources(expires_at) WHERE expires_at IS NOT NULL`,\n // emulate cascade for the virtual table\n `CREATE TRIGGER trg_documents_delete_vec\nAFTER DELETE ON documents\nBEGIN\n DELETE FROM vec_chunks\n WHERE source_id = OLD.source_id\n AND document_id = OLD.id;\nEND`,\n];\n"],
|
|
5
|
-
"mappings": "AAKO,MAAM,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAYd,SAAS,QAAQ,WAAmB;AACzC,SAAO;AAAA;AAAA;AAAA,oBAGW,SAAS;AAAA;AAAA;AAG7B;AAEO,MAAM,aAAa;AAAA,EACxB;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAEA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAOF;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../../../src/lib/stores/sqlite/node-sqlite.ts"],
|
|
4
|
-
"sourcesContent": ["import { open } from './open.js';\nimport { SQLiteStore } from './sqlite.js';\n\nexport function nodeSQLite(dbName: string, dimension: number) {\n return new SQLiteStore(open(dbName), dimension);\n}\n"],
|
|
5
|
-
"mappings": "AAAA,SAAS,YAAY;AACrB,SAAS,mBAAmB;AAErB,SAAS,WAAW,QAAgB,WAAmB;AAC5D,SAAO,IAAI,YAAY,KAAK,MAAM,GAAG,SAAS;AAChD;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
|
@@ -1,11 +0,0 @@
|
|
|
1
|
-
import { DatabaseSync } from "node:sqlite";
|
|
2
|
-
import * as sqliteVec from "sqlite-vec";
|
|
3
|
-
function open(path) {
|
|
4
|
-
const db = new DatabaseSync(path, { allowExtension: true });
|
|
5
|
-
db.loadExtension(sqliteVec.getLoadablePath());
|
|
6
|
-
return db;
|
|
7
|
-
}
|
|
8
|
-
export {
|
|
9
|
-
open
|
|
10
|
-
};
|
|
11
|
-
//# sourceMappingURL=open.js.map
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../../../src/lib/stores/sqlite/open.ts"],
|
|
4
|
-
"sourcesContent": ["import { DatabaseSync } from 'node:sqlite';\nimport * as sqliteVec from 'sqlite-vec';\n\n// The store's statements use the vec0 module, so every connection that runs\n// them, the dev database of the declaration included, loads sqlite-vec.\nexport function open(path: string) {\n const db = new DatabaseSync(path, { allowExtension: true });\n db.loadExtension(sqliteVec.getLoadablePath());\n return db;\n}\n"],
|
|
5
|
-
"mappings": "AAAA,SAAS,oBAAoB;AAC7B,YAAY,eAAe;AAIpB,SAAS,KAAK,MAAc;AACjC,QAAM,KAAK,IAAI,aAAa,MAAM,EAAE,gBAAgB,KAAK,CAAC;AAC1D,KAAG,cAAc,UAAU,gBAAgB,CAAC;AAC5C,SAAO;AACT;",
|
|
6
|
-
"names": []
|
|
7
|
-
}
|
|
@@ -1,245 +0,0 @@
|
|
|
1
|
-
import { Declaration } from "@zukhruf/sqlite-drift";
|
|
2
|
-
import { transaction } from "@zukhruf/sqlite-transaction";
|
|
3
|
-
import dedent from "dedent";
|
|
4
|
-
import { SETUP, statements, vectors } from "./ddl.sqlite.js";
|
|
5
|
-
import { open } from "./open.js";
|
|
6
|
-
const DEFAULT_TOP_N = 10;
|
|
7
|
-
const tables = new Declaration(statements, { rows: "keep" });
|
|
8
|
-
class SQLiteStore {
|
|
9
|
-
#db;
|
|
10
|
-
constructor(db, dimension) {
|
|
11
|
-
this.#db = db;
|
|
12
|
-
this.#db.exec(SETUP);
|
|
13
|
-
const chunks = new Declaration([vectors(dimension)], {
|
|
14
|
-
rows: "replace",
|
|
15
|
-
devDatabase: () => open(":memory:")
|
|
16
|
-
});
|
|
17
|
-
transaction(
|
|
18
|
-
this.#db,
|
|
19
|
-
() => {
|
|
20
|
-
const stale = chunks.drift(this.#db) === "stale";
|
|
21
|
-
chunks.apply(this.#db);
|
|
22
|
-
if (stale) this.#db.exec("DELETE FROM sources");
|
|
23
|
-
},
|
|
24
|
-
{ mode: "immediate" }
|
|
25
|
-
);
|
|
26
|
-
tables.apply(this.#db);
|
|
27
|
-
}
|
|
28
|
-
#searchByDocument(inputs) {
|
|
29
|
-
const stmt = this.#db.prepare(dedent`
|
|
30
|
-
SELECT v.content, v.distance, v.document_id, d.metadata
|
|
31
|
-
FROM vec_chunks v
|
|
32
|
-
JOIN documents d ON d.id = v.document_id
|
|
33
|
-
WHERE v.source_id = ?
|
|
34
|
-
AND v.document_id = ?
|
|
35
|
-
AND v.embedding MATCH vec_normalize(vec_f32(?))
|
|
36
|
-
AND v.k = ?
|
|
37
|
-
ORDER BY v.distance ASC
|
|
38
|
-
`);
|
|
39
|
-
return stmt.all(
|
|
40
|
-
inputs.sourceId,
|
|
41
|
-
inputs.documentId,
|
|
42
|
-
inputs.embedding,
|
|
43
|
-
inputs.k
|
|
44
|
-
);
|
|
45
|
-
}
|
|
46
|
-
#searchBySource(inputs) {
|
|
47
|
-
const stmt = this.#db.prepare(dedent`
|
|
48
|
-
SELECT v.content, v.distance, v.document_id, d.metadata
|
|
49
|
-
FROM vec_chunks v
|
|
50
|
-
JOIN documents d ON d.id = v.document_id
|
|
51
|
-
WHERE v.source_id = ?
|
|
52
|
-
AND v.embedding MATCH vec_normalize(vec_f32(?))
|
|
53
|
-
AND v.k = ?
|
|
54
|
-
ORDER BY v.distance ASC
|
|
55
|
-
`);
|
|
56
|
-
return stmt.all(inputs.sourceId, inputs.embedding, inputs.k);
|
|
57
|
-
}
|
|
58
|
-
async search(query, options, embedder) {
|
|
59
|
-
const { embeddings } = await embedder([query]);
|
|
60
|
-
if (!embeddings.length) {
|
|
61
|
-
return [];
|
|
62
|
-
}
|
|
63
|
-
const vectorBlob = vectorToBlob(embeddings[0]);
|
|
64
|
-
const topN = options.topN;
|
|
65
|
-
if (options.documentId) {
|
|
66
|
-
const rows2 = this.#searchByDocument({
|
|
67
|
-
sourceId: options.sourceId,
|
|
68
|
-
documentId: options.documentId,
|
|
69
|
-
embedding: vectorBlob,
|
|
70
|
-
k: topN ?? DEFAULT_TOP_N
|
|
71
|
-
});
|
|
72
|
-
return rows2.map(toSearchResult);
|
|
73
|
-
}
|
|
74
|
-
const rows = this.#searchBySource({
|
|
75
|
-
sourceId: options.sourceId,
|
|
76
|
-
embedding: vectorBlob,
|
|
77
|
-
k: topN ?? DEFAULT_TOP_N
|
|
78
|
-
});
|
|
79
|
-
return rows.map(toSearchResult);
|
|
80
|
-
}
|
|
81
|
-
#upsertSource(inputs) {
|
|
82
|
-
const stmt = this.#db.prepare(dedent`
|
|
83
|
-
INSERT INTO sources (source_id) VALUES (?)
|
|
84
|
-
ON CONFLICT(source_id) DO UPDATE SET updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')
|
|
85
|
-
`);
|
|
86
|
-
return stmt.run(inputs.sourceId);
|
|
87
|
-
}
|
|
88
|
-
#upsertSourceWithExpiry(inputs) {
|
|
89
|
-
const stmt = this.#db.prepare(dedent`
|
|
90
|
-
INSERT INTO sources (source_id, expires_at) VALUES (?, ?)
|
|
91
|
-
ON CONFLICT(source_id) DO UPDATE SET
|
|
92
|
-
updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now'),
|
|
93
|
-
expires_at=excluded.expires_at
|
|
94
|
-
`);
|
|
95
|
-
return stmt.run(inputs.sourceId, inputs.expiresAt);
|
|
96
|
-
}
|
|
97
|
-
upsertDoc(inputs) {
|
|
98
|
-
const stmt = this.#db.prepare(dedent`
|
|
99
|
-
INSERT INTO documents (id, source_id, cid, metadata)
|
|
100
|
-
VALUES (?, ?, ?, json(?))
|
|
101
|
-
ON CONFLICT(id) DO UPDATE SET
|
|
102
|
-
cid=excluded.cid,
|
|
103
|
-
metadata=COALESCE(excluded.metadata, documents.metadata),
|
|
104
|
-
updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')
|
|
105
|
-
WHERE documents.cid != excluded.cid;
|
|
106
|
-
`);
|
|
107
|
-
return stmt.run(
|
|
108
|
-
inputs.documentId,
|
|
109
|
-
inputs.sourceId,
|
|
110
|
-
inputs.cid,
|
|
111
|
-
inputs.metadata ? JSON.stringify(inputs.metadata) : null
|
|
112
|
-
);
|
|
113
|
-
}
|
|
114
|
-
insertDoc(inputs) {
|
|
115
|
-
const stmt = this.#db.prepare(dedent`
|
|
116
|
-
INSERT INTO vec_chunks (source_id, document_id, content, embedding)
|
|
117
|
-
VALUES (?, ?, ?, vec_normalize(vec_f32(?)))
|
|
118
|
-
`);
|
|
119
|
-
return (chunk) => {
|
|
120
|
-
stmt.run(
|
|
121
|
-
inputs.sourceId,
|
|
122
|
-
inputs.documentId,
|
|
123
|
-
chunk.content,
|
|
124
|
-
vectorToBlob(chunk.embedding)
|
|
125
|
-
);
|
|
126
|
-
};
|
|
127
|
-
}
|
|
128
|
-
delete(inputs) {
|
|
129
|
-
const stmt = this.#db.prepare(dedent`
|
|
130
|
-
DELETE FROM vec_chunks WHERE source_id = ? AND document_id = ?
|
|
131
|
-
`);
|
|
132
|
-
return stmt.run(inputs.sourceId, inputs.documentId);
|
|
133
|
-
}
|
|
134
|
-
sourceExists(sourceId) {
|
|
135
|
-
const stmt = this.#db.prepare(dedent`
|
|
136
|
-
SELECT 1 FROM sources WHERE source_id = ? LIMIT 1
|
|
137
|
-
`);
|
|
138
|
-
const row = stmt.get(sourceId);
|
|
139
|
-
return Boolean(row);
|
|
140
|
-
}
|
|
141
|
-
sourceExpired(sourceId) {
|
|
142
|
-
const stmt = this.#db.prepare(dedent`
|
|
143
|
-
SELECT 1 FROM sources
|
|
144
|
-
WHERE source_id = ?
|
|
145
|
-
AND expires_at IS NOT NULL
|
|
146
|
-
AND expires_at <= strftime('%Y-%m-%dT%H:%M:%fZ','now')
|
|
147
|
-
LIMIT 1
|
|
148
|
-
`);
|
|
149
|
-
const row = stmt.get(sourceId);
|
|
150
|
-
return Boolean(row);
|
|
151
|
-
}
|
|
152
|
-
setSourceExpiry(sourceId, expiryDate) {
|
|
153
|
-
const stmt = this.#db.prepare(dedent`
|
|
154
|
-
UPDATE sources
|
|
155
|
-
SET expires_at = ?, updated_at = strftime('%Y-%m-%dT%H:%M:%fZ','now')
|
|
156
|
-
WHERE source_id = ?
|
|
157
|
-
`);
|
|
158
|
-
stmt.run(expiryDate.toISOString(), sourceId);
|
|
159
|
-
}
|
|
160
|
-
async index(sourceId, corpus, expiryDate) {
|
|
161
|
-
if (expiryDate) {
|
|
162
|
-
this.#upsertSourceWithExpiry({
|
|
163
|
-
expiresAt: expiryDate.toISOString(),
|
|
164
|
-
sourceId
|
|
165
|
-
});
|
|
166
|
-
} else {
|
|
167
|
-
this.#upsertSource({ sourceId });
|
|
168
|
-
}
|
|
169
|
-
const info = this.upsertDoc({
|
|
170
|
-
documentId: corpus.id,
|
|
171
|
-
sourceId,
|
|
172
|
-
cid: corpus.cid,
|
|
173
|
-
metadata: corpus.metadata
|
|
174
|
-
});
|
|
175
|
-
const changed = info.changes > 0;
|
|
176
|
-
if (!changed) {
|
|
177
|
-
return;
|
|
178
|
-
}
|
|
179
|
-
const insert = this.insertDoc({ sourceId, documentId: corpus.id });
|
|
180
|
-
transaction(
|
|
181
|
-
this.#db,
|
|
182
|
-
() => {
|
|
183
|
-
this.delete({ sourceId, documentId: corpus.id });
|
|
184
|
-
},
|
|
185
|
-
{ mode: "immediate" }
|
|
186
|
-
);
|
|
187
|
-
const batchSize = 32;
|
|
188
|
-
let batch = [];
|
|
189
|
-
const flush = () => {
|
|
190
|
-
if (!batch.length) return;
|
|
191
|
-
transaction(
|
|
192
|
-
this.#db,
|
|
193
|
-
() => {
|
|
194
|
-
for (let i = 0; i < batch.length; i++) {
|
|
195
|
-
insert(batch[i]);
|
|
196
|
-
}
|
|
197
|
-
},
|
|
198
|
-
{ mode: "immediate" }
|
|
199
|
-
);
|
|
200
|
-
batch = [];
|
|
201
|
-
};
|
|
202
|
-
for await (const chunk of corpus.chunker()) {
|
|
203
|
-
batch.push(chunk);
|
|
204
|
-
if (batch.length >= batchSize) flush();
|
|
205
|
-
}
|
|
206
|
-
flush();
|
|
207
|
-
}
|
|
208
|
-
}
|
|
209
|
-
function vectorToBlob(vector) {
|
|
210
|
-
if (vector instanceof Float32Array) {
|
|
211
|
-
const copied = new Float32Array(vector.length);
|
|
212
|
-
copied.set(vector);
|
|
213
|
-
return Buffer.from(copied.buffer);
|
|
214
|
-
}
|
|
215
|
-
const floatArray = new Float32Array(vector);
|
|
216
|
-
return Buffer.from(floatArray.buffer);
|
|
217
|
-
}
|
|
218
|
-
function toSearchResult(row) {
|
|
219
|
-
if (typeof row !== "object" || row === null || !("content" in row) || typeof row.content !== "string" || !("distance" in row) || typeof row.distance !== "number" || !("document_id" in row) || typeof row.document_id !== "string" || !("metadata" in row) || row.metadata !== null && typeof row.metadata !== "string") {
|
|
220
|
-
throw new Error("Unexpected vector search row");
|
|
221
|
-
}
|
|
222
|
-
return {
|
|
223
|
-
content: row.content,
|
|
224
|
-
distance: row.distance,
|
|
225
|
-
document_id: row.document_id,
|
|
226
|
-
metadata: parseMetadata(row.metadata)
|
|
227
|
-
};
|
|
228
|
-
}
|
|
229
|
-
function parseMetadata(value) {
|
|
230
|
-
if (value === null) return null;
|
|
231
|
-
try {
|
|
232
|
-
const parsed = JSON.parse(value);
|
|
233
|
-
return isRecord(parsed) ? parsed : null;
|
|
234
|
-
} catch {
|
|
235
|
-
return null;
|
|
236
|
-
}
|
|
237
|
-
}
|
|
238
|
-
function isRecord(value) {
|
|
239
|
-
return typeof value === "object" && value !== null && !Array.isArray(value);
|
|
240
|
-
}
|
|
241
|
-
export {
|
|
242
|
-
SQLiteStore,
|
|
243
|
-
vectorToBlob
|
|
244
|
-
};
|
|
245
|
-
//# sourceMappingURL=sqlite.js.map
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 3,
|
|
3
|
-
"sources": ["../../../../src/lib/stores/sqlite/sqlite.ts"],
|
|
4
|
-
"sourcesContent": ["import { Declaration } from '@zukhruf/sqlite-drift';\nimport { transaction } from '@zukhruf/sqlite-transaction';\nimport dedent from 'dedent';\nimport type { DatabaseSync } from 'node:sqlite';\n\nimport type {\n Chunk,\n Corpus,\n Embedder,\n SearchOptions,\n SearchResult,\n Store,\n} from '../store.js';\nimport { SETUP, statements, vectors } from './ddl.sqlite.js';\nimport { open } from './open.js';\n\nconst DEFAULT_TOP_N = 10;\n\nconst tables = new Declaration(statements, { rows: 'keep' });\n\nexport class SQLiteStore implements Store {\n readonly #db: DatabaseSync;\n constructor(db: DatabaseSync, dimension: number) {\n this.#db = db;\n this.#db.exec(SETUP);\n const chunks = new Declaration([vectors(dimension)], {\n rows: 'replace',\n devDatabase: () => open(':memory:'),\n });\n transaction(\n this.#db,\n () => {\n const stale = chunks.drift(this.#db) === 'stale';\n chunks.apply(this.#db);\n // A replaced vec_chunks is empty, yet upsertDoc skips a document whose\n // cid is unchanged and similaritySearch skips a source that exists.\n // Deleting the sources cascades to their documents, so the next\n // ingest embeds every document again.\n if (stale) this.#db.exec('DELETE FROM sources');\n },\n { mode: 'immediate' },\n );\n tables.apply(this.#db);\n }\n\n #searchByDocument(inputs: {\n sourceId: string;\n documentId: string;\n embedding: Buffer<ArrayBufferLike>;\n k: number;\n }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tSELECT v.content, v.distance, v.document_id, d.metadata\n\t\t\t\tFROM vec_chunks v\n\t\t\t\tJOIN documents d ON d.id = v.document_id\n\t\t\t\tWHERE v.source_id = ?\n\t\t\t\t\tAND v.document_id = ?\n\t\t\t\t\tAND v.embedding MATCH vec_normalize(vec_f32(?))\n\t\t\t\t\tAND v.k = ?\n\t\t\t\tORDER BY v.distance ASC\n\t\t\t`);\n return stmt.all(\n inputs.sourceId,\n inputs.documentId,\n inputs.embedding,\n inputs.k,\n );\n }\n #searchBySource(inputs: {\n sourceId: string;\n embedding: Buffer<ArrayBufferLike>;\n k: number;\n }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tSELECT v.content, v.distance, v.document_id, d.metadata\n\t\t\t\tFROM vec_chunks v\n\t\t\t\tJOIN documents d ON d.id = v.document_id\n\t\t\t\tWHERE v.source_id = ?\n\t\t\t\t\tAND v.embedding MATCH vec_normalize(vec_f32(?))\n\t\t\t\t\tAND v.k = ?\n\t\t\t\tORDER BY v.distance ASC\n\t\t\t`);\n return stmt.all(inputs.sourceId, inputs.embedding, inputs.k);\n }\n async search(query: string, options: SearchOptions, embedder: Embedder) {\n const { embeddings } = await embedder([query]);\n if (!embeddings.length) {\n return [];\n }\n const vectorBlob = vectorToBlob(embeddings[0]);\n const topN = options.topN;\n\n if (options.documentId) {\n const rows = this.#searchByDocument({\n sourceId: options.sourceId,\n documentId: options.documentId,\n embedding: vectorBlob,\n k: topN ?? DEFAULT_TOP_N,\n });\n return rows.map(toSearchResult);\n }\n\n const rows = this.#searchBySource({\n sourceId: options.sourceId,\n embedding: vectorBlob,\n k: topN ?? DEFAULT_TOP_N,\n });\n return rows.map(toSearchResult);\n }\n #upsertSource(inputs: { sourceId: string }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tINSERT INTO sources (source_id) VALUES (?)\n\t\t\t\tON CONFLICT(source_id) DO UPDATE SET updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')\n\t\t\t`);\n return stmt.run(inputs.sourceId);\n }\n #upsertSourceWithExpiry(inputs: { sourceId: string; expiresAt: string }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tINSERT INTO sources (source_id, expires_at) VALUES (?, ?)\n\t\t\t\tON CONFLICT(source_id) DO UPDATE SET\n\t\t\t\t\tupdated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now'),\n\t\t\t\t\texpires_at=excluded.expires_at\n\t\t\t`);\n return stmt.run(inputs.sourceId, inputs.expiresAt);\n }\n upsertDoc(inputs: {\n documentId: string;\n sourceId: string;\n cid: string;\n metadata?: Record<string, any>;\n }) {\n const stmt = this.#db.prepare(dedent`\n INSERT INTO documents (id, source_id, cid, metadata)\n VALUES (?, ?, ?, json(?))\n ON CONFLICT(id) DO UPDATE SET\n cid=excluded.cid,\n metadata=COALESCE(excluded.metadata, documents.metadata),\n updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')\n WHERE documents.cid != excluded.cid;\n `);\n return stmt.run(\n inputs.documentId,\n inputs.sourceId,\n inputs.cid,\n inputs.metadata ? JSON.stringify(inputs.metadata) : null,\n );\n }\n insertDoc(inputs: { sourceId: string; documentId: string }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tINSERT INTO vec_chunks (source_id, document_id, content, embedding)\n\t\t\t\tVALUES (?, ?, ?, vec_normalize(vec_f32(?)))\n\t\t\t`);\n return (chunk: Chunk) => {\n stmt.run(\n inputs.sourceId,\n inputs.documentId,\n chunk.content,\n vectorToBlob(chunk.embedding),\n );\n };\n }\n delete(inputs: { sourceId: string; documentId: string }) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tDELETE FROM vec_chunks WHERE source_id = ? AND document_id = ?\n\t\t\t`);\n return stmt.run(inputs.sourceId, inputs.documentId);\n }\n sourceExists(sourceId: string) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tSELECT 1 FROM sources WHERE source_id = ? LIMIT 1\n\t\t\t`);\n const row = stmt.get(sourceId);\n return Boolean(row);\n }\n sourceExpired(sourceId: string) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tSELECT 1 FROM sources\n\t\t\t\tWHERE source_id = ?\n\t\t\t\t\tAND expires_at IS NOT NULL\n\t\t\t\t\tAND expires_at <= strftime('%Y-%m-%dT%H:%M:%fZ','now')\n\t\t\t\tLIMIT 1\n\t\t\t`);\n const row = stmt.get(sourceId);\n return Boolean(row);\n }\n setSourceExpiry(sourceId: string, expiryDate: Date) {\n const stmt = this.#db.prepare(dedent`\n\t\t\t\tUPDATE sources\n\t\t\t\tSET expires_at = ?, updated_at = strftime('%Y-%m-%dT%H:%M:%fZ','now')\n\t\t\t\tWHERE source_id = ?\n\t\t\t`);\n stmt.run(expiryDate.toISOString(), sourceId);\n }\n async index(sourceId: string, corpus: Corpus, expiryDate?: Date) {\n if (expiryDate) {\n this.#upsertSourceWithExpiry({\n expiresAt: expiryDate.toISOString(),\n sourceId,\n });\n } else {\n this.#upsertSource({ sourceId });\n }\n\n const info = this.upsertDoc({\n documentId: corpus.id,\n sourceId,\n cid: corpus.cid,\n metadata: corpus.metadata,\n });\n const changed = info.changes > 0;\n\n if (!changed) {\n return;\n }\n\n const insert = this.insertDoc({ sourceId, documentId: corpus.id });\n // Delete previous rows once before inserting\n transaction(\n this.#db,\n () => {\n this.delete({ sourceId, documentId: corpus.id });\n },\n { mode: 'immediate' },\n );\n\n const batchSize = 32;\n let batch: Chunk[] = [];\n const flush = () => {\n if (!batch.length) return;\n transaction(\n this.#db,\n () => {\n for (let i = 0; i < batch.length; i++) {\n insert(batch[i]);\n }\n },\n { mode: 'immediate' },\n );\n batch = [];\n };\n\n for await (const chunk of corpus.chunker()) {\n batch.push(chunk);\n if (batch.length >= batchSize) flush();\n }\n // flush any remaining\n flush();\n }\n}\n\nexport function vectorToBlob(vector: number[] | Float32Array): Buffer {\n if (vector instanceof Float32Array) {\n // Copy into a fresh Buffer to avoid retaining references to a larger batch tensor buffer\n const copied = new Float32Array(vector.length);\n copied.set(vector);\n return Buffer.from(copied.buffer);\n }\n const floatArray = new Float32Array(vector);\n return Buffer.from(floatArray.buffer);\n}\n\nfunction toSearchResult(row: unknown): SearchResult {\n if (\n typeof row !== 'object' ||\n row === null ||\n !('content' in row) ||\n typeof row.content !== 'string' ||\n !('distance' in row) ||\n typeof row.distance !== 'number' ||\n !('document_id' in row) ||\n typeof row.document_id !== 'string' ||\n !('metadata' in row) ||\n (row.metadata !== null && typeof row.metadata !== 'string')\n ) {\n throw new Error('Unexpected vector search row');\n }\n return {\n content: row.content,\n distance: row.distance,\n document_id: row.document_id,\n metadata: parseMetadata(row.metadata),\n };\n}\n\nfunction parseMetadata(value: string | null): Record<string, unknown> | null {\n if (value === null) return null;\n try {\n const parsed: unknown = JSON.parse(value);\n return isRecord(parsed) ? parsed : null;\n } catch {\n return null;\n }\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n return typeof value === 'object' && value !== null && !Array.isArray(value);\n}\n"],
|
|
5
|
-
"mappings": "AAAA,SAAS,mBAAmB;AAC5B,SAAS,mBAAmB;AAC5B,OAAO,YAAY;AAWnB,SAAS,OAAO,YAAY,eAAe;AAC3C,SAAS,YAAY;AAErB,MAAM,gBAAgB;AAEtB,MAAM,SAAS,IAAI,YAAY,YAAY,EAAE,MAAM,OAAO,CAAC;AAEpD,MAAM,YAA6B;AAAA,EAC/B;AAAA,EACT,YAAY,IAAkB,WAAmB;AAC/C,SAAK,MAAM;AACX,SAAK,IAAI,KAAK,KAAK;AACnB,UAAM,SAAS,IAAI,YAAY,CAAC,QAAQ,SAAS,CAAC,GAAG;AAAA,MACnD,MAAM;AAAA,MACN,aAAa,MAAM,KAAK,UAAU;AAAA,IACpC,CAAC;AACD;AAAA,MACE,KAAK;AAAA,MACL,MAAM;AACJ,cAAM,QAAQ,OAAO,MAAM,KAAK,GAAG,MAAM;AACzC,eAAO,MAAM,KAAK,GAAG;AAKrB,YAAI,MAAO,MAAK,IAAI,KAAK,qBAAqB;AAAA,MAChD;AAAA,MACA,EAAE,MAAM,YAAY;AAAA,IACtB;AACA,WAAO,MAAM,KAAK,GAAG;AAAA,EACvB;AAAA,EAEA,kBAAkB,QAKf;AACD,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAS9B;AACA,WAAO,KAAK;AAAA,MACV,OAAO;AAAA,MACP,OAAO;AAAA,MACP,OAAO;AAAA,MACP,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA,gBAAgB,QAIb;AACD,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAQ9B;AACA,WAAO,KAAK,IAAI,OAAO,UAAU,OAAO,WAAW,OAAO,CAAC;AAAA,EAC7D;AAAA,EACA,MAAM,OAAO,OAAe,SAAwB,UAAoB;AACtE,UAAM,EAAE,WAAW,IAAI,MAAM,SAAS,CAAC,KAAK,CAAC;AAC7C,QAAI,CAAC,WAAW,QAAQ;AACtB,aAAO,CAAC;AAAA,IACV;AACA,UAAM,aAAa,aAAa,WAAW,CAAC,CAAC;AAC7C,UAAM,OAAO,QAAQ;AAErB,QAAI,QAAQ,YAAY;AACtB,YAAMA,QAAO,KAAK,kBAAkB;AAAA,QAClC,UAAU,QAAQ;AAAA,QAClB,YAAY,QAAQ;AAAA,QACpB,WAAW;AAAA,QACX,GAAG,QAAQ;AAAA,MACb,CAAC;AACD,aAAOA,MAAK,IAAI,cAAc;AAAA,IAChC;AAEA,UAAM,OAAO,KAAK,gBAAgB;AAAA,MAChC,UAAU,QAAQ;AAAA,MAClB,WAAW;AAAA,MACX,GAAG,QAAQ;AAAA,IACb,CAAC;AACD,WAAO,KAAK,IAAI,cAAc;AAAA,EAChC;AAAA,EACA,cAAc,QAA8B;AAC1C,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA,IAG9B;AACA,WAAO,KAAK,IAAI,OAAO,QAAQ;AAAA,EACjC;AAAA,EACA,wBAAwB,QAAiD;AACvE,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA,IAK9B;AACA,WAAO,KAAK,IAAI,OAAO,UAAU,OAAO,SAAS;AAAA,EACnD;AAAA,EACA,UAAU,QAKP;AACD,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,OAQ3B;AACH,WAAO,KAAK;AAAA,MACV,OAAO;AAAA,MACP,OAAO;AAAA,MACP,OAAO;AAAA,MACP,OAAO,WAAW,KAAK,UAAU,OAAO,QAAQ,IAAI;AAAA,IACtD;AAAA,EACF;AAAA,EACA,UAAU,QAAkD;AAC1D,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA,IAG9B;AACA,WAAO,CAAC,UAAiB;AACvB,WAAK;AAAA,QACH,OAAO;AAAA,QACP,OAAO;AAAA,QACP,MAAM;AAAA,QACN,aAAa,MAAM,SAAS;AAAA,MAC9B;AAAA,IACF;AAAA,EACF;AAAA,EACA,OAAO,QAAkD;AACvD,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA,IAE9B;AACA,WAAO,KAAK,IAAI,OAAO,UAAU,OAAO,UAAU;AAAA,EACpD;AAAA,EACA,aAAa,UAAkB;AAC7B,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA,IAE9B;AACA,UAAM,MAAM,KAAK,IAAI,QAAQ;AAC7B,WAAO,QAAQ,GAAG;AAAA,EACpB;AAAA,EACA,cAAc,UAAkB;AAC9B,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAM9B;AACA,UAAM,MAAM,KAAK,IAAI,QAAQ;AAC7B,WAAO,QAAQ,GAAG;AAAA,EACpB;AAAA,EACA,gBAAgB,UAAkB,YAAkB;AAClD,UAAM,OAAO,KAAK,IAAI,QAAQ;AAAA;AAAA;AAAA;AAAA,IAI9B;AACA,SAAK,IAAI,WAAW,YAAY,GAAG,QAAQ;AAAA,EAC7C;AAAA,EACA,MAAM,MAAM,UAAkB,QAAgB,YAAmB;AAC/D,QAAI,YAAY;AACd,WAAK,wBAAwB;AAAA,QAC3B,WAAW,WAAW,YAAY;AAAA,QAClC;AAAA,MACF,CAAC;AAAA,IACH,OAAO;AACL,WAAK,cAAc,EAAE,SAAS,CAAC;AAAA,IACjC;AAEA,UAAM,OAAO,KAAK,UAAU;AAAA,MAC1B,YAAY,OAAO;AAAA,MACnB;AAAA,MACA,KAAK,OAAO;AAAA,MACZ,UAAU,OAAO;AAAA,IACnB,CAAC;AACD,UAAM,UAAU,KAAK,UAAU;AAE/B,QAAI,CAAC,SAAS;AACZ;AAAA,IACF;AAEA,UAAM,SAAS,KAAK,UAAU,EAAE,UAAU,YAAY,OAAO,GAAG,CAAC;AAEjE;AAAA,MACE,KAAK;AAAA,MACL,MAAM;AACJ,aAAK,OAAO,EAAE,UAAU,YAAY,OAAO,GAAG,CAAC;AAAA,MACjD;AAAA,MACA,EAAE,MAAM,YAAY;AAAA,IACtB;AAEA,UAAM,YAAY;AAClB,QAAI,QAAiB,CAAC;AACtB,UAAM,QAAQ,MAAM;AAClB,UAAI,CAAC,MAAM,OAAQ;AACnB;AAAA,QACE,KAAK;AAAA,QACL,MAAM;AACJ,mBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,mBAAO,MAAM,CAAC,CAAC;AAAA,UACjB;AAAA,QACF;AAAA,QACA,EAAE,MAAM,YAAY;AAAA,MACtB;AACA,cAAQ,CAAC;AAAA,IACX;AAEA,qBAAiB,SAAS,OAAO,QAAQ,GAAG;AAC1C,YAAM,KAAK,KAAK;AAChB,UAAI,MAAM,UAAU,UAAW,OAAM;AAAA,IACvC;AAEA,UAAM;AAAA,EACR;AACF;AAEO,SAAS,aAAa,QAAyC;AACpE,MAAI,kBAAkB,cAAc;AAElC,UAAM,SAAS,IAAI,aAAa,OAAO,MAAM;AAC7C,WAAO,IAAI,MAAM;AACjB,WAAO,OAAO,KAAK,OAAO,MAAM;AAAA,EAClC;AACA,QAAM,aAAa,IAAI,aAAa,MAAM;AAC1C,SAAO,OAAO,KAAK,WAAW,MAAM;AACtC;AAEA,SAAS,eAAe,KAA4B;AAClD,MACE,OAAO,QAAQ,YACf,QAAQ,QACR,EAAE,aAAa,QACf,OAAO,IAAI,YAAY,YACvB,EAAE,cAAc,QAChB,OAAO,IAAI,aAAa,YACxB,EAAE,iBAAiB,QACnB,OAAO,IAAI,gBAAgB,YAC3B,EAAE,cAAc,QACf,IAAI,aAAa,QAAQ,OAAO,IAAI,aAAa,UAClD;AACA,UAAM,IAAI,MAAM,8BAA8B;AAAA,EAChD;AACA,SAAO;AAAA,IACL,SAAS,IAAI;AAAA,IACb,UAAU,IAAI;AAAA,IACd,aAAa,IAAI;AAAA,IACjB,UAAU,cAAc,IAAI,QAAQ;AAAA,EACtC;AACF;AAEA,SAAS,cAAc,OAAsD;AAC3E,MAAI,UAAU,KAAM,QAAO;AAC3B,MAAI;AACF,UAAM,SAAkB,KAAK,MAAM,KAAK;AACxC,WAAO,SAAS,MAAM,IAAI,SAAS;AAAA,EACrC,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,SAAS,OAAkD;AAClE,SAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;",
|
|
6
|
-
"names": ["rows"]
|
|
7
|
-
}
|
package/dist/lib/stores/store.js
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
//# sourceMappingURL=store.js.map
|