@dxos/assistant 0.9.0 → 0.9.1-staging.ee54ba693a

package/dist/lib/neutral/extraction/index.mjs

@@ -1,78 +1,120 @@
 import "../chunk-J5LGTIGS.mjs";
 
-// src/extraction/extraction.ts
-import * as Schema from "effect/Schema";
-import { Operation } from "@dxos/compute";
-import { Obj, Type } from "@dxos/echo";
-import { DXN } from "@dxos/keys";
-import { log } from "@dxos/log";
+// src/extraction/extraction-llm-function/proper-noun-extraction.ts
+import * as LanguageModel from "@effect/ai/LanguageModel";
+import * as Effect from "effect/Effect";
+import * as Schema2 from "effect/Schema";
+import { AiService } from "@dxos/ai";
+import { Obj } from "@dxos/echo";
 import { Message } from "@dxos/types";
-var __dxlog_file = "/__w/dxos/dxos/packages/core/compute/assistant/src/extraction/extraction.ts";
-var ExtractionInput = Schema.Struct({
-  message: Type.getSchema(Message.Message),
-  objects: Schema.optional(Schema.Array(Obj.Unknown)),
-  options: Schema.optional(Schema.Struct({
-    timeout: Schema.optional(Schema.Number),
-    fallbackToRaw: Schema.optional(Schema.Boolean)
-  }))
-});
-var ExtractionOutput = Schema.Struct({
-  message: Type.getSchema(Message.Message),
-  timeElapsed: Schema.Number
-});
-var extractionAnthropicFunction = Operation.make({
-  meta: {
-    key: DXN.make("org.dxos.function.extraction.anthropic"),
-    name: "Entity Extraction (Anthropic)",
-    description: "Extract entities from transcript using Anthropic LLM.",
-    icon: "ph--text-t--regular"
-  },
-  input: ExtractionInput,
-  output: ExtractionOutput
-});
-var extractionNerFunction = Operation.make({
-  meta: {
-    key: DXN.make("org.dxos.function.extraction.ner"),
-    name: "Entity Extraction (NER)",
-    description: "Extract entities from transcript using Named Entity Recognition.",
-    icon: "ph--text-t--regular"
-  },
-  input: ExtractionInput,
-  output: ExtractionOutput
+
+// src/extraction/quotes.ts
+import * as Schema from "effect/Schema";
+import { Filter, Query } from "@dxos/echo";
+import { EID, EntityId } from "@dxos/keys";
+var ReferencedQuotes = Schema.Struct({
+  references: Schema.Array(Schema.Struct({
+    id: Schema.String,
+    quote: Schema.String
+  })).pipe(Schema.mutable).annotations({
+    // TODO(burdon): Does this description make sense?
+    description: `
+      The references to the context objects that are mentioned in the transcript. 
+      Quote should match the original transcript text exactly, while id is the id of the context object.
+    `
+  })
 });
-var processTranscriptMessage = async (params) => {
-  try {
-    if (params.options?.timeout && params.options.timeout > 0) {
-    } else {
-    }
-  } catch (error) {
-    if (params.options?.fallbackToRaw) {
-      log.warn("failed to process message, falling back to raw text", {
-        error
-      }, { "~LogMeta": "~LogMeta", F: __dxlog_file, L: 60, S: void 0 });
-      return {
-        message: params.input.message,
-        timeElapsed: 0
-      };
-    } else {
-      throw error;
-    }
-  }
+var findReferences = async (nouns, db, { searchKind = "full-text" } = {}) => {
+  const normalizedNouns = Array.from(new Set(nouns.map((noun) => noun.trim()).filter((noun) => noun.length > 0))).sort((leftNoun, rightNoun) => rightNoun.length - leftNoun.length);
+  const references = await Promise.all(normalizedNouns.map(async (noun) => {
+    const objects = await db.query(Query.select(Filter.text(noun, {
+      type: searchKind
+    }))).run();
+    return objects.length > 0 ? [
+      {
+        id: objects[0].id.toString(),
+        quote: noun
+      }
+    ] : [];
+  }));
   return {
-    message: params.input.message,
-    timeElapsed: 0
+    references: references.flat()
   };
 };
+var insertReferences = (text, quotes) => {
+  const references = quotes.references.filter((quote) => EntityId.isValid(quote.id)).sort((left, right) => right.quote.length - left.quote.length);
+  for (const quote of references) {
+    const regex = new RegExp(quote.quote.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"), "gi");
+    text = replaceOutsideLinks(text, regex, `[${quote.quote}](${EID.make({
+      entityId: quote.id
+    })})`);
+  }
+  return text;
+};
+var MARKDOWN_LINK = /\[[^\]]*\]\([^)]*\)/g;
+var replaceOutsideLinks = (text, regex, replacement) => {
+  let result = "";
+  let plainStart = 0;
+  for (const link of text.matchAll(MARKDOWN_LINK)) {
+    const index = link.index ?? 0;
+    result += text.slice(plainStart, index).replace(regex, replacement) + link[0];
+    plainStart = index + link[0].length;
+  }
+  return result + text.slice(plainStart).replace(regex, replacement);
+};
+
+// src/extraction/extraction-llm-function/proper-noun-extraction.ts
+var EXTRACTION_MODEL = "com.anthropic.model.claude-haiku-4-5.default";
+var ProperNouns = Schema2.Struct({
+  properNouns: Schema2.Array(Schema2.String).annotations({
+    description: "Proper nouns (names of people, organizations, places, products) mentioned in the text."
+  })
+});
+var extractProperNouns = (text) => Effect.gen(function* () {
+  const { value } = yield* Effect.scoped(LanguageModel.generateObject({
+    schema: ProperNouns,
+    prompt: [
+      "Extract every proper noun (people, organizations, places, products) mentioned in the transcript text below.",
+      "Return only the surface strings exactly as they appear. Exclude common nouns and pronouns.",
+      "",
+      "Transcript:",
+      text
+    ].join("\n")
+  }));
+  return value.properNouns.map((noun) => noun.trim()).filter((noun) => noun.length >= 3);
+}).pipe(Effect.provide(AiService.model(EXTRACTION_MODEL)));
+var enrichTranscriptMessage = (message, { db }) => Effect.gen(function* () {
+  const blocks = yield* Effect.forEach(message.blocks, (block) => Effect.gen(function* () {
+    if (block._tag !== "transcript") {
+      return block;
+    }
+    const nouns = yield* extractProperNouns(block.text);
+    const references = yield* Effect.promise(() => findReferences([
+      ...nouns
+    ], db));
+    return {
+      ...block,
+      text: insertReferences(block.text, references)
+    };
+  }));
+  return Obj.make(Message.Message, {
+    id: message.id,
+    sender: message.sender,
+    created: message.created,
+    properties: message.properties,
+    blocks
+  });
+});
 
 // src/extraction/extraction-ner-function/named-entity-recognition.ts
 import { pipeline } from "@xenova/transformers";
-import { log as log2 } from "@dxos/log";
-var __dxlog_file2 = "/__w/dxos/dxos/packages/core/compute/assistant/src/extraction/extraction-ner-function/named-entity-recognition.ts";
+import { log } from "@dxos/log";
+var __dxlog_file = "/__w/dxos/dxos/packages/core/compute/assistant/src/extraction/extraction-ner-function/named-entity-recognition.ts";
 var _ner;
 var getNer = () => {
   if (!_ner) {
     _ner = pipeline("ner", "Xenova/bert-base-NER").then((ner) => {
-      log2.info("NER model is ready", void 0, { "~LogMeta": "~LogMeta", F: __dxlog_file2, L: 14, S: void 0 });
+      log.info("NER model is ready", void 0, { "~LogMeta": "~LogMeta", F: __dxlog_file, L: 14, S: void 0 });
       return ner;
     });
   }
@@ -140,69 +182,16 @@ var combineNerTokens = (group) => {
   }
   return combinedToken;
 };
-
-// src/extraction/quotes.ts
-import * as Schema2 from "effect/Schema";
-import { Filter, Query } from "@dxos/echo";
-import { EID, EntityId } from "@dxos/keys";
-import { log as log3 } from "@dxos/log";
-var __dxlog_file3 = "/__w/dxos/dxos/packages/core/compute/assistant/src/extraction/quotes.ts";
-var ReferencedQuotes = Schema2.Struct({
-  references: Schema2.Array(Schema2.Struct({
-    id: Schema2.String,
-    quote: Schema2.String
-  })).pipe(Schema2.mutable).annotations({
-    // TODO(burdon): Does this description make sense?
-    description: `
-      The references to the context objects that are mentioned in the transcript. 
-      Quote should match the original transcript text exactly, while id is the id of the context object.
-    `
-  })
-});
-var findQuotes = async (quotes, db) => {
-  const references = await Promise.all(quotes.map(async (quote) => {
-    const objects = await db.query(Query.select(Filter.text(quote, {
-      type: "vector"
-    }))).run();
-    log3.info("found quotes in vector index", {
-      objects,
-      quote
-    }, { "~LogMeta": "~LogMeta", F: __dxlog_file3, L: 25, S: void 0 });
-    return objects.length > 0 ? [
-      {
-        id: objects[0].id.toString(),
-        quote
-      }
-    ] : [];
-  }));
-  return {
-    references: references.flat()
-  };
-};
-var insertReferences = (text, quotes) => {
-  for (const quote of quotes.references) {
-    if (!EntityId.isValid(quote.id)) {
-      continue;
-    }
-    const regex = new RegExp(quote.quote.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"), "gi");
-    text = text.replace(regex, `[${quote.quote}](${EID.make({
-      entityId: quote.id
-    })})`);
-  }
-  return text;
-};
 export {
-  ExtractionInput,
-  ExtractionOutput,
+  ProperNouns,
   ReferencedQuotes,
   combineNerTokens,
   createTokenGroups,
+  enrichTranscriptMessage,
   extractFullEntities,
-  extractionAnthropicFunction,
-  extractionNerFunction,
-  findQuotes,
+  extractProperNouns,
+  findReferences,
   getNer,
-  insertReferences,
-  processTranscriptMessage
+  insertReferences
 };
 //# sourceMappingURL=index.mjs.map

package/dist/lib/neutral/extraction/index.mjs.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
-  "sources": ["../../../../src/extraction/extraction.ts", "../../../../src/extraction/extraction-ner-function/named-entity-recognition.ts", "../../../../src/extraction/quotes.ts"],
-  "sourcesContent": ["//\n// Copyright 2025 DXOS.org\n//\n\nimport * as Schema from 'effect/Schema';\n\nimport { Operation } from '@dxos/compute';\nimport { Obj, Type } from '@dxos/echo';\nimport { DXN } from '@dxos/keys';\nimport { log } from '@dxos/log';\nimport { Message } from '@dxos/types';\n\nexport const ExtractionInput = Schema.Struct({\n  message: Type.getSchema(Message.Message),\n  objects: Schema.optional(Schema.Array(Obj.Unknown)),\n  options: Schema.optional(\n    Schema.Struct({\n      timeout: Schema.optional(Schema.Number),\n      fallbackToRaw: Schema.optional(Schema.Boolean),\n    }),\n  ),\n});\nexport type ExtractionInput = Schema.Schema.Type<typeof ExtractionInput>;\nexport const ExtractionOutput = Schema.Struct({\n  message: Type.getSchema(Message.Message),\n  timeElapsed: Schema.Number,\n});\nexport type ExtractionOutput = Schema.Schema.Type<typeof ExtractionOutput>;\nexport type ProcessTranscriptMessageProps = {\n  input: ExtractionInput;\n  function: Operation.Definition<ExtractionInput, ExtractionOutput>;\n\n  options?: {\n    /**\n     * Timeout for the entity extraction.\n     */\n    timeout?: number;\n\n    /**\n     * Fallback to raw text if the entity extraction fails.\n     * Otherwise the function will throw an error.\n     * @default false\n     */\n    fallbackToRaw?: boolean;\n  };\n};\n\nexport type ExtractionFunction = Operation.Definition<ExtractionInput, ExtractionOutput>;\n\n/**\n * Placeholder for Anthropic-based entity extraction.\n * Implementation is currently disabled pending restoration of extraction logic.\n */\nexport const extractionAnthropicFunction = Operation.make({\n  meta: {\n    key: DXN.make('org.dxos.function.extraction.anthropic'),\n    name: 'Entity Extraction (Anthropic)',\n    description: 'Extract entities from transcript using Anthropic LLM.',\n    icon: 'ph--text-t--regular',\n  },\n  input: ExtractionInput,\n  output: ExtractionOutput,\n});\n\n/**\n * Placeholder for NER-based entity extraction.\n * Implementation is currently disabled pending restoration of extraction logic.\n */\nexport const extractionNerFunction = Operation.make({\n  meta: {\n    key: DXN.make('org.dxos.function.extraction.ner'),\n    name: 'Entity Extraction (NER)',\n    description: 'Extract entities from transcript using Named Entity Recognition.',\n    icon: 'ph--text-t--regular',\n  },\n  input: ExtractionInput,\n  output: ExtractionOutput,\n});\n\n/**\n * Extract entities from the transcript message and add them to the message.\n */\n// TODO(dmaretskyi): Move context to a vector search index.\nexport const processTranscriptMessage = async (params: ProcessTranscriptMessageProps): Promise<ExtractionOutput> => {\n  try {\n    if (params.options?.timeout && params.options.timeout > 0) {\n      // return await asyncTimeout(params.executor.invoke(params.function, params.input), params.options.timeout);\n    } else {\n      // return await params.executor.invoke(params.function, params.input);\n    }\n  } catch (error) {\n    if (params.options?.fallbackToRaw) {\n      log.warn('failed to process message, falling back to raw text', { error });\n      return {\n        message: params.input.message,\n        timeElapsed: 0,\n      };\n    } else {\n      throw error;\n    }\n  }\n\n  // TODO(dmaretskyi): Restore extraction logic.\n  return { message: params.input.message, timeElapsed: 0 };\n};\n", "//\n// Copyright 2025 DXOS.org\n//\n\nimport {\n  type TokenClassificationOutput,\n  type TokenClassificationPipelineType,\n  type TokenClassificationSingle,\n  pipeline,\n} from '@xenova/transformers';\n\nimport { log } from '@dxos/log';\n\nlet _ner: Promise<TokenClassificationPipelineType>;\n\n/**\n * Named Entity Recognition pipeline.\n * Initializes the pipeline on first call.\n * @returns The singleton promise that resolves to a token classification pipeline.\n */\nexport const getNer = () => {\n  if (!_ner) {\n    _ner = pipeline('ner', 'Xenova/bert-base-NER').then((ner) => {\n      log.info('NER model is ready');\n      return ner;\n    });\n  }\n\n  return _ner;\n};\n\n/**\n * Extracts entities from text using the named entity recognition pipeline and does simple grouping of tokens.\n */\nexport const extractFullEntities = async (text: string): Promise<TokenClassificationSingle[]> => {\n  const ner = await getNer();\n  const entities = (await ner(text)) as TokenClassificationOutput;\n  const tokenGroups = createTokenGroups(entities);\n  const combinedEntities = tokenGroups.map(combineNerTokens);\n  return combinedEntities;\n};\n\nexport const createTokenGroups = (tokens: TokenClassificationSingle[]): TokenClassificationSingle[][] => {\n  const tokenGroups: TokenClassificationSingle[][] = [];\n\n  const canBeCombined = (prevToken: TokenClassificationSingle, currentToken: TokenClassificationSingle) => {\n    const [currentState, currentEntity] = currentToken.entity.split('-'); // B, I, O\n    const [prevState, prevEntity] = prevToken.entity.split('-'); // B, I, O\n    if (prevEntity !== currentEntity) {\n      return false;\n    }\n    if (prevState === 'O') {\n      return false;\n    }\n    if (currentState === 'I') {\n      return true;\n    }\n    if ((prevState === 'B' || prevState === 'I') && currentState === 'B') {\n      return currentToken.word.startsWith('#');\n    }\n    return false;\n  };\n\n  for (const token of tokens) {\n    if (tokenGroups.length === 0) {\n      tokenGroups.push([token]);\n      continue;\n    }\n    const lastTokenGroup = tokenGroups.at(-1)!;\n    if (canBeCombined(lastTokenGroup.at(-1)!, token)) {\n      lastTokenGroup.push(token);\n      continue;\n    }\n    tokenGroups.push([token]);\n  }\n\n  return tokenGroups;\n};\n\nexport const combineNerTokens = (group: TokenClassificationSingle[]): TokenClassificationSingle => {\n  const combinedToken: TokenClassificationSingle = {\n    entity: group.at(0)!.entity.split('-')[1],\n    score: group.reduce((acc, token) => acc + token.score, 0) / group.length,\n    index: group.at(0)!.index,\n    word: group.at(0)!.word,\n    start: group.at(0)!.start,\n    end: group.at(-1)!.end,\n  };\n\n  for (const token of group.slice(1)) {\n    if (token.word.startsWith('#')) {\n      combinedToken.word += token.word.replace(/^#+/, '');\n    } else {\n      combinedToken.word += ' ' + token.word;\n    }\n  }\n\n  return combinedToken;\n};\n", "//\n// Copyright 2025 DXOS.org\n//\n\nimport * as Schema from 'effect/Schema';\n\nimport { type Database, Filter, Query } from '@dxos/echo';\nimport { EID, EntityId } from '@dxos/keys';\nimport { log } from '@dxos/log';\n\nexport const ReferencedQuotes = Schema.Struct({\n  references: Schema.Array(\n    Schema.Struct({\n      id: Schema.String,\n      quote: Schema.String, // TODO(burdon): Quote?\n    }),\n  )\n    .pipe(Schema.mutable)\n    .annotations({\n      // TODO(burdon): Does this description make sense?\n      description: `\n      The references to the context objects that are mentioned in the transcript. \n      Quote should match the original transcript text exactly, while id is the id of the context object.\n    `,\n    }),\n});\nexport type ReferencedQuotes = Schema.Schema.Type<typeof ReferencedQuotes>;\n\nexport const findQuotes = async (quotes: string[], db: Database.Database): Promise<ReferencedQuotes> => {\n  const references = await Promise.all(\n    quotes.map(async (quote) => {\n      const objects = await db.query(Query.select(Filter.text(quote, { type: 'vector' }))).run();\n      log.info('found quotes in vector index', { objects, quote });\n      return objects.length > 0 ? [{ id: objects[0].id.toString(), quote }] : [];\n    }),\n  );\n  return { references: references.flat() };\n};\n\n/**\n * Finds and replaces all quotes with DXNs references.\n */\n// TODO(dmaretskyi): Lookup and verifiy ids from provided context.\nexport const insertReferences = (text: string, quotes: ReferencedQuotes) => {\n  for (const quote of quotes.references) {\n    if (!EntityId.isValid(quote.id)) {\n      continue;\n    }\n\n    // Use a case-insensitive regular expression to replace the quote.\n    const regex = new RegExp(quote.quote.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&'), 'gi');\n    text = text.replace(regex, `[${quote.quote}](${EID.make({ entityId: quote.id })})`);\n  }\n\n  return text;\n};\n"],
-  "mappings": ";;;AAIA,YAAYA,YAAY;AAExB,SAASC,iBAAiB;AAC1B,SAASC,KAAKC,YAAY;AAC1B,SAASC,WAAW;AACpB,SAASC,WAAW;AACpB,SAASC,eAAe;AAExB,IAAA,eAAaC;AAEXC,IAASR,kBAAuBS,cAAMP;EACtCQ,SAASV,KAAAA,UAAe,QACfW,OAAO;WACHX,gBAAgBA,aAAOY,IAAM,OAAA,CAAA;WACtCC,gBAAsBC,cAASd;IACjC,SAAA,gBAAA,aAAA;IAED,eAAA,gBAAA,cAAA;EAEH,CAAA,CAAA;;AAEEe,IAAAA,mBAA0B,cAAA;EACzB,SAAA,KAAA,UAAA,QAAA,OAAA;EAuBH,aAAA;;AAMIC,IAASC,8BAAK,UAAA,KAAA;QACdC;IACAC,KAAAA,IAAAA,KAAa,wCAAA;IACbC,MAAM;IACR,aAAA;IACAC,MAAOd;EACPe;EACC,OAAA;EAEH,QAAA;;AAMIN,IAASC,wBAAK,UAAA,KAAA;QACdC;IACAC,KAAAA,IAAAA,KAAa,kCAAA;IACbC,MAAM;IACR,aAAA;IACAC,MAAOd;EACPe;EACC,OAAA;EAEH,QAAA;;IAMQC,2BAA2BA,OAAOb,WAAQc;MAC5C;AACF,QAAA,OAAO,SAAA,WAAA,OAAA,QAAA,UAAA,GAAA;IAEP,OAAA;IAEA;WACEnB,OAAQ;eAA0DoB,SAAAA,eAAAA;AAAM,UAAA,KAAA,uDAAA;QACxE;uBACWF,YAAY,GAACG,cAAO,GAAA,IAAA,GAAA,OAAA,CAAA;aAC7BX;QACF,SAAA,OAAA,MAAA;QACK,aAAA;MACL;IACF,OAAA;AACF,YAAA;IAEA;EACA;SAAwCA;IAAe,SAAA,OAAA,MAAA;IACvD,aAAA;;;;;ACpGF,SAIEY,gBACK;AAEP,SAASC,OAAAA,YAAW;AAEpB,IAAIC,gBAAAA;;AASAA,IAAOF,SAAS,MAAA;aACdC;WACA,SAAOE,OAAAA,sBAAAA,EAAAA,KAAAA,CAAAA,QAAAA;AACT,MAAAF,KAAA,KAAA,sBAAA,QAAA,EAAA,YAAA,YAAA,GAAAG,eAAA,GAAA,IAAA,GAAA,OAAA,CAAA;AACF,aAAA;IAEA,CAAA;EACA;AAEF,SAAA;;AAKQC,IAAAA,sBAAsBC,OAAAA,SAAAA;AAC5B,QAAMC,MAAAA,MAAAA,OAAcC;AACpB,QAAMC,WAAAA,MAAAA,IAAmBF,IAAAA;AACzB,QAAA,cAAOE,kBAAAA,QAAAA;AACP,QAAA,mBAAA,YAAA,IAAA,gBAAA;AAEF,SAAO;;AAGL,IAAMC,oBAAiBC,CAAAA,WAAsCC;QAC3D,cAAOC,CAAAA;QACP,gBAAkBC,CAAAA,WAAW,iBAAaC;AAC1C,UAAID,CAAAA,cAAeE,aAAe,IAAA,aAAA,OAAA,MAAA,GAAA;UAChC,CAAA,WAAO,UAAA,IAAA,UAAA,OAAA,MAAA,GAAA;AACT,QAAA,eAAA,eAAA;AACIC,aAAAA;;AAEJ,QAAA,cAAA,KAAA;AACIJ,aAAAA;;AAEJ,QAAA,iBAAA,KAAA;AACKI,aAAAA;;AAEL,SAAA,cAAA,OAAA,cAAA,QAAA,iBAAA,KAAA;AACA,aAAO,aAAA,KAAA,WAAA,GAAA;IACT;AAEA,WAAWC;;aAEPX,SAAYY,QAAK;oBAACD,WAAAA,GAAAA;kBAAM,KAAA;QACxB;MACF,CAAA;AACA;IACA;UACEE,iBAAmB,YAACF,GAAAA,EAAAA;QACpB,cAAA,eAAA,GAAA,EAAA,GAAA,KAAA,GAAA;AACF,qBAAA,KAAA,KAAA;AACAX;;gBAAwB,KAAA;MAC1B;IAEA,CAAA;EACA;AAEF,SAAO;;IAEHQ,mBAAqBA,CAAAA,UAAa;QAClCM,gBAAaC;IACbC,QAAOC,MAAMC,GAAG,CAAA,EAAIF,OAAK,MAAA,GAAA,EAAA,CAAA;IACzBG,OAAMF,MAAMC,OAAOC,CAAAA,KAAI,UAAA,MAAA,MAAA,OAAA,CAAA,IAAA,MAAA;IACvBC,OAAOH,MAAMC,GAAG,CAAA,EAAIE;IACpBC,MAAKJ,MAAMC,GAAI,CAAA,EAAIG;IACrB,OAAA,MAAA,GAAA,CAAA,EAAA;IAEA,KAAK,MAAMV,GAAAA,EAASM,EAAAA;;aAEhBK,SAAAA,MAAkB,MAAIX,CAAAA,GAAAA;AACxB,QAAA,MAAO,KAAA,WAAA,GAAA,GAAA;AACLW,oBAAcH,QAAQ,MAAMR,KAAAA,QAAU,OAAA,EAAA;IACxC,OAAA;AACF,oBAAA,QAAA,MAAA,MAAA;IAEA;EACA;;;;;AC9FF,YAAYY,aAAY;AAExB,SAAwBC,QAAQC,aAAa;AAC7C,SAASC,KAAKC,gBAAgB;AAC9B,SAASC,OAAAA,YAAW;AAEpB,IAAAC,gBAAaC;IAGHP,mBAAa,eAAA;cACHQ,cAAM,eAAA;IAGrBC,IAAKT;IAEJ,OAAA;MACAU,KAAAA,eAAc,EAAA,YAAA;;;;;;EAQpB,CAAA;;IAGM,aAAgB,OAASC,QAAMT,OAAMU;qBAAkC,MAAA,QAAA,IAAA,OAAA,IAAA,OAAA,UAAA;AAAS,UAAKC,UAAG,MAAA,GAAA,MAAA,MAAA,OAAA,OAAA,KAAA,OAAA;MACpFC,MAAK;SAAkCC,IAAAA;SAASC,KAAAA,gCAAAA;MAAM;MAC1D;qBAA6B,YAAA,GAAAV,eAAA,GAAA,IAAA,GAAA,OAAA,CAAA;mBAAMS,SAAWE,IAAGC;;QAAkB,IAAA,QAAA,CAAA,EAAA,GAAA,SAAA;QAAO;MAC5E;IAEF,IAAO,CAAA;;AAAgC,SAAA;IACvC,YAAA,WAAA,KAAA;EAEF;;IAMSd,mBAAiBY,CAAAA,MAAQ,WAAG;aAC/B,SAAA,OAAA,YAAA;AACF,QAAA,CAAA,SAAA,QAAA,MAAA,EAAA,GAAA;AAEA;IACA;UAC0DG,QAAUH,IAAAA,OAAQ,MAAA,MAAA,QAAA,uBAAA,MAAA,GAAA,IAAA;AAAC,WAAK,KAAA,QAAA,OAAA,IAAA,MAAA,KAAA,KAAA,IAAA,KAAA;MACpF,UAAA,MAAA;IAEA,CAAA,CAAA,GAAOI;EACP;;;",
-  "names": ["Schema", "Operation", "Obj", "Type", "DXN", "log", "Message", "ExtractionInput", "objects", "Array", "options", "Struct", "Number", "fallbackToRaw", "optional", "timeElapsed", "key", "make", "name", "description", "icon", "input", "output", "params", "timeout", "error", "message", "pipeline", "log", "_ner", "ner", "__dxlog_file", "entities", "text", "tokenGroups", "createTokenGroups", "combinedEntities", "canBeCombined", "prevToken", "currentToken", "currentState", "prevEntity", "entity", "currentEntity", "prevState", "token", "push", "lastTokenGroup", "score", "reduce", "index", "group", "at", "word", "start", "end", "combinedToken", "Schema", "Filter", "Query", "EID", "EntityId", "log", "__dxlog_file", "ReferencedQuotes", "String", "pipe", "description", "query", "select", "run", "info", "objects", "quote", "id", "toString", "entityId", "text"]
+  "sources": ["../../../../src/extraction/extraction-llm-function/proper-noun-extraction.ts", "../../../../src/extraction/quotes.ts", "../../../../src/extraction/extraction-ner-function/named-entity-recognition.ts"],
+  "sourcesContent": ["//\n// Copyright 2025 DXOS.org\n//\n\nimport * as LanguageModel from '@effect/ai/LanguageModel';\nimport * as Effect from 'effect/Effect';\nimport * as Schema from 'effect/Schema';\n\nimport { AiService } from '@dxos/ai';\nimport { type Database, Obj } from '@dxos/echo';\nimport { Message } from '@dxos/types';\n\nimport { findReferences, insertReferences } from '../quotes';\n\nconst EXTRACTION_MODEL = 'com.anthropic.model.claude-haiku-4-5.default';\n\n/**\n * Proper nouns extracted from transcript text.\n */\nexport const ProperNouns = Schema.Struct({\n  properNouns: Schema.Array(Schema.String).annotations({\n    description: 'Proper nouns (names of people, organizations, places, products) mentioned in the text.',\n  }),\n});\nexport type ProperNouns = Schema.Schema.Type<typeof ProperNouns>;\n\n/**\n * Extracts proper nouns from transcript text using a small LLM (Haiku). Provides the LanguageModel\n * internally so the residual requirement is just {@link AiService.AiService}.\n */\nexport const extractProperNouns = (text: string) =>\n  Effect.gen(function* () {\n    const { value } = yield* Effect.scoped(\n      LanguageModel.generateObject({\n        schema: ProperNouns,\n        prompt: [\n          'Extract every proper noun (people, organizations, places, products) mentioned in the transcript text below.',\n          'Return only the surface strings exactly as they appear. Exclude common nouns and pronouns.',\n          '',\n          'Transcript:',\n          text,\n        ].join('\\n'),\n      }),\n    );\n\n    // Drop short tokens (e.g. \"IT\"): insertReferences replaces case-insensitive substrings, so a\n    // short noun would match inside unrelated words (\"secur[IT]y\") and corrupt neighbouring links.\n    return value.properNouns.map((noun) => noun.trim()).filter((noun) => noun.length >= 3);\n  }).pipe(Effect.provide(AiService.model(EXTRACTION_MODEL)));\n\n/**\n * Enriches a transcript message: for each transcript block, extracts proper nouns, links them to\n * objects in the space via full-text search, and rewrites the text with `[noun](dxn)` references.\n * Requires {@link AiService.AiService}; takes the space database handle.\n */\nexport const enrichTranscriptMessage = (message: Message.Message, { db }: { db: Database.Database }) =>\n  Effect.gen(function* () {\n    const blocks = yield* Effect.forEach(message.blocks, (block) =>\n      Effect.gen(function* () {\n        if (block._tag !== 'transcript') {\n          return block;\n        }\n\n        const nouns = yield* extractProperNouns(block.text);\n        const references = yield* Effect.promise(() => findReferences([...nouns], db));\n        return { ...block, text: insertReferences(block.text, references) };\n      }),\n    );\n\n    // Preserve the message identity and metadata; enrichment only rewrites transcript block text.\n    return Obj.make(Message.Message, {\n      id: message.id,\n      sender: message.sender,\n      created: message.created,\n      properties: message.properties,\n      blocks,\n    });\n  });\n", "//\n// Copyright 2025 DXOS.org\n//\n\nimport * as Schema from 'effect/Schema';\n\nimport { type Database, Filter, Query } from '@dxos/echo';\nimport { EID, EntityId } from '@dxos/keys';\n\nexport const ReferencedQuotes = Schema.Struct({\n  references: Schema.Array(\n    Schema.Struct({\n      id: Schema.String,\n      quote: Schema.String, // TODO(burdon): Quote?\n    }),\n  )\n    .pipe(Schema.mutable)\n    .annotations({\n      // TODO(burdon): Does this description make sense?\n      description: `\n      The references to the context objects that are mentioned in the transcript. \n      Quote should match the original transcript text exactly, while id is the id of the context object.\n    `,\n    }),\n});\nexport type ReferencedQuotes = Schema.Schema.Type<typeof ReferencedQuotes>;\n\nexport type FindReferencesOptions = {\n  /** Search backend. Full-text matches proper nouns precisely; vector matches semantically. */\n  searchKind?: 'full-text' | 'vector';\n};\n\n/**\n * Finds the best-matching context object for each noun and returns it as a reference.\n */\nexport const findReferences = async (\n  nouns: string[],\n  db: Database.Database,\n  { searchKind = 'full-text' }: FindReferencesOptions = {},\n): Promise<ReferencedQuotes> => {\n  // Deduplicate and process longer nouns first: `insertReferences` rewrites case-insensitive\n  // substrings, so a repeated or shorter overlapping noun would re-replace already-linked text and\n  // produce nested/broken markdown links.\n  const normalizedNouns = Array.from(new Set(nouns.map((noun) => noun.trim()).filter((noun) => noun.length > 0))).sort(\n    (leftNoun, rightNoun) => rightNoun.length - leftNoun.length,\n  );\n\n  const references = await Promise.all(\n    normalizedNouns.map(async (noun) => {\n      const objects = await db.query(Query.select(Filter.text(noun, { type: searchKind }))).run();\n      return objects.length > 0 ? [{ id: objects[0].id.toString(), quote: noun }] : [];\n    }),\n  );\n  return { references: references.flat() };\n};\n\n/**\n * Finds and replaces all quotes with DXNs references.\n */\n// TODO(dmaretskyi): Lookup and verifiy ids from provided context.\nexport const insertReferences = (text: string, quotes: ReferencedQuotes): string => {\n  // Process longer quotes first so e.g. \"Sarah Johnson\" is linked before \"Sarah\". Combined with the\n  // link-aware replacement below this stops a shorter or overlapping quote from rewriting text that\n  // is already part of an inserted link (which would produce nested `[[..](..)](..)` markdown).\n  const references = quotes.references\n    .filter((quote) => EntityId.isValid(quote.id))\n    .sort((left, right) => right.quote.length - left.quote.length);\n\n  for (const quote of references) {\n    // Case-insensitive match of the (regex-escaped) quote.\n    const regex = new RegExp(quote.quote.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&'), 'gi');\n    text = replaceOutsideLinks(text, regex, `[${quote.quote}](${EID.make({ entityId: quote.id })})`);\n  }\n\n  return text;\n};\n\nconst MARKDOWN_LINK = /\\[[^\\]]*\\]\\([^)]*\\)/g;\n\n/** Applies a replacement to every span of `text` that is not already inside a `[label](url)` link. */\nconst replaceOutsideLinks = (text: string, regex: RegExp, replacement: string): string => {\n  let result = '';\n  let plainStart = 0;\n  for (const link of text.matchAll(MARKDOWN_LINK)) {\n    const index = link.index ?? 0;\n    result += text.slice(plainStart, index).replace(regex, replacement) + link[0];\n    plainStart = index + link[0].length;\n  }\n  return result + text.slice(plainStart).replace(regex, replacement);\n};\n", "//\n// Copyright 2025 DXOS.org\n//\n\nimport {\n  type TokenClassificationOutput,\n  type TokenClassificationPipelineType,\n  type TokenClassificationSingle,\n  pipeline,\n} from '@xenova/transformers';\n\nimport { log } from '@dxos/log';\n\nlet _ner: Promise<TokenClassificationPipelineType>;\n\n/**\n * Named Entity Recognition pipeline.\n * Initializes the pipeline on first call.\n * @returns The singleton promise that resolves to a token classification pipeline.\n */\nexport const getNer = () => {\n  if (!_ner) {\n    _ner = pipeline('ner', 'Xenova/bert-base-NER').then((ner) => {\n      log.info('NER model is ready');\n      return ner;\n    });\n  }\n\n  return _ner;\n};\n\n/**\n * Extracts entities from text using the named entity recognition pipeline and does simple grouping of tokens.\n */\nexport const extractFullEntities = async (text: string): Promise<TokenClassificationSingle[]> => {\n  const ner = await getNer();\n  const entities = (await ner(text)) as TokenClassificationOutput;\n  const tokenGroups = createTokenGroups(entities);\n  const combinedEntities = tokenGroups.map(combineNerTokens);\n  return combinedEntities;\n};\n\nexport const createTokenGroups = (tokens: TokenClassificationSingle[]): TokenClassificationSingle[][] => {\n  const tokenGroups: TokenClassificationSingle[][] = [];\n\n  const canBeCombined = (prevToken: TokenClassificationSingle, currentToken: TokenClassificationSingle) => {\n    const [currentState, currentEntity] = currentToken.entity.split('-'); // B, I, O\n    const [prevState, prevEntity] = prevToken.entity.split('-'); // B, I, O\n    if (prevEntity !== currentEntity) {\n      return false;\n    }\n    if (prevState === 'O') {\n      return false;\n    }\n    if (currentState === 'I') {\n      return true;\n    }\n    if ((prevState === 'B' || prevState === 'I') && currentState === 'B') {\n      return currentToken.word.startsWith('#');\n    }\n    return false;\n  };\n\n  for (const token of tokens) {\n    if (tokenGroups.length === 0) {\n      tokenGroups.push([token]);\n      continue;\n    }\n    const lastTokenGroup = tokenGroups.at(-1)!;\n    if (canBeCombined(lastTokenGroup.at(-1)!, token)) {\n      lastTokenGroup.push(token);\n      continue;\n    }\n    tokenGroups.push([token]);\n  }\n\n  return tokenGroups;\n};\n\nexport const combineNerTokens = (group: TokenClassificationSingle[]): TokenClassificationSingle => {\n  const combinedToken: TokenClassificationSingle = {\n    entity: group.at(0)!.entity.split('-')[1],\n    score: group.reduce((acc, token) => acc + token.score, 0) / group.length,\n    index: group.at(0)!.index,\n    word: group.at(0)!.word,\n    start: group.at(0)!.start,\n    end: group.at(-1)!.end,\n  };\n\n  for (const token of group.slice(1)) {\n    if (token.word.startsWith('#')) {\n      combinedToken.word += token.word.replace(/^#+/, '');\n    } else {\n      combinedToken.word += ' ' + token.word;\n    }\n  }\n\n  return combinedToken;\n};\n"],
+  "mappings": ";;;AAIA,YAAYA,mBAAmB;AAC/B,YAAYC,YAAY;AACxB,YAAYC,aAAY;AAExB,SAASC,iBAAiB;AAC1B,SAAwBC,WAAW;AACnC,SAASC,eAAe;;;ACNxB,YAAYC,YAAY;AAExB,SAAwBC,QAAQC,aAAa;AAC7C,SAASC,KAAKC,gBAAgB;AAEvB,IAAMC,mBAA0BC,cAAO;EAC5CC,YAAmBC,aACVF,cAAO;IACZG,IAAWC;IACXC,OAAcD;EAChB,CAAA,CAAA,EAECE,KAAYC,cAAO,EACnBC,YAAY;;IAEXC,aAAa;;;;EAIf,CAAA;AACJ,CAAA;AAWO,IAAMC,iBAAiB,OAC5BC,OACAC,IACA,EAAEC,aAAa,YAAW,IAA4B,CAAC,MAAC;AAKxD,QAAMC,kBAAkBZ,MAAMa,KAAK,IAAIC,IAAIL,MAAMM,IAAI,CAACC,SAASA,KAAKC,KAAI,CAAA,EAAIC,OAAO,CAACF,SAASA,KAAKG,SAAS,CAAA,CAAA,CAAA,EAAKC,KAC9G,CAACC,UAAUC,cAAcA,UAAUH,SAASE,SAASF,MAAM;AAG7D,QAAMpB,aAAa,MAAMwB,QAAQC,IAC/BZ,gBAAgBG,IAAI,OAAOC,SAAAA;AACzB,UAAMS,UAAU,MAAMf,GAAGgB,MAAMhC,MAAMiC,OAAOlC,OAAOmC,KAAKZ,MAAM;MAAEa,MAAMlB;IAAW,CAAA,CAAA,CAAA,EAAKmB,IAAG;AACzF,WAAOL,QAAQN,SAAS,IAAI;MAAC;QAAElB,IAAIwB,QAAQ,CAAA,EAAGxB,GAAG8B,SAAQ;QAAI5B,OAAOa;MAAK;QAAK,CAAA;EAChF,CAAA,CAAA;AAEF,SAAO;IAAEjB,YAAYA,WAAWiC,KAAI;EAAG;AACzC;AAMO,IAAMC,mBAAmB,CAACL,MAAcM,WAAAA;AAI7C,QAAMnC,aAAamC,OAAOnC,WACvBmB,OAAO,CAACf,UAAUP,SAASuC,QAAQhC,MAAMF,EAAE,CAAA,EAC3CmB,KAAK,CAACgB,MAAMC,UAAUA,MAAMlC,MAAMgB,SAASiB,KAAKjC,MAAMgB,MAAM;AAE/D,aAAWhB,SAASJ,YAAY;AAE9B,UAAMuC,QAAQ,IAAIC,OAAOpC,MAAMA,MAAMqC,QAAQ,uBAAuB,MAAA,GAAS,IAAA;AAC7EZ,WAAOa,oBAAoBb,MAAMU,OAAO,IAAInC,MAAMA,KAAK,KAAKR,IAAI+C,KAAK;MAAEC,UAAUxC,MAAMF;IAAG,CAAA,CAAA,GAAK;EACjG;AAEA,SAAO2B;AACT;AAEA,IAAMgB,gBAAgB;AAGtB,IAAMH,sBAAsB,CAACb,MAAcU,OAAeO,gBAAAA;AACxD,MAAIC,SAAS;AACb,MAAIC,aAAa;AACjB,aAAWC,QAAQpB,KAAKqB,SAASL,aAAAA,GAAgB;AAC/C,UAAMM,QAAQF,KAAKE,SAAS;AAC5BJ,cAAUlB,KAAKuB,MAAMJ,YAAYG,KAAAA,EAAOV,QAAQF,OAAOO,WAAAA,IAAeG,KAAK,CAAA;AAC3ED,iBAAaG,QAAQF,KAAK,CAAA,EAAG7B;EAC/B;AACA,SAAO2B,SAASlB,KAAKuB,MAAMJ,UAAAA,EAAYP,QAAQF,OAAOO,WAAAA;AACxD;;;AD3EA,IAAMO,mBAAmB;AAKlB,IAAMC,cAAqBC,eAAO;EACvCC,aAAoBC,cAAaC,cAAM,EAAEC,YAAY;IACnDC,aAAa;EACf,CAAA;AACF,CAAA;AAOO,IAAMC,qBAAqB,CAACC,SAC1BC,WAAI,aAAA;AACT,QAAM,EAAEC,MAAK,IAAK,OAAcC,cAChBC,6BAAe;IAC3BC,QAAQb;IACRc,QAAQ;MACN;MACA;MACA;MACA;MACAN;MACAO,KAAK,IAAA;EACT,CAAA,CAAA;AAKF,SAAOL,MAAMR,YAAYc,IAAI,CAACC,SAASA,KAAKC,KAAI,CAAA,EAAIC,OAAO,CAACF,SAASA,KAAKG,UAAU,CAAA;AACtF,CAAA,EAAGC,KAAYC,eAAQC,UAAUC,MAAMzB,gBAAAA,CAAAA,CAAAA;AAOlC,IAAM0B,0BAA0B,CAACC,SAA0B,EAAEC,GAAE,MAC7DlB,WAAI,aAAA;AACT,QAAMmB,SAAS,OAAcC,eAAQH,QAAQE,QAAQ,CAACE,UAC7CrB,WAAI,aAAA;AACT,QAAIqB,MAAMC,SAAS,cAAc;AAC/B,aAAOD;IACT;AAEA,UAAME,QAAQ,OAAOzB,mBAAmBuB,MAAMtB,IAAI;AAClD,UAAMyB,aAAa,OAAcC,eAAQ,MAAMC,eAAe;SAAIH;OAAQL,EAAAA,CAAAA;AAC1E,WAAO;MAAE,GAAGG;MAAOtB,MAAM4B,iBAAiBN,MAAMtB,MAAMyB,UAAAA;IAAY;EACpE,CAAA,CAAA;AAIF,SAAOI,IAAIC,KAAKC,QAAQA,SAAS;IAC/BC,IAAId,QAAQc;IACZC,QAAQf,QAAQe;IAChBC,SAAShB,QAAQgB;IACjBC,YAAYjB,QAAQiB;IACpBf;EACF,CAAA;AACF,CAAA;;;AEzEF,SAIEgB,gBACK;AAEP,SAASC,WAAW;AAEpB,IAAIC,eAAAA;;AASAA,IAAOF,SAAS,MAAA;aACdC;WACA,SAAOE,OAAAA,sBAAAA,EAAAA,KAAAA,CAAAA,QAAAA;AACT,UAAA,KAAA,sBAAA,QAAA,EAAA,YAAA,YAAA,GAAA,cAAA,GAAA,IAAA,GAAA,OAAA,CAAA;AACF,aAAA;IAEA,CAAA;EACA;AAEF,SAAA;;AAKQC,IAAAA,sBAAsBC,OAAAA,SAAAA;AAC5B,QAAMC,MAAAA,MAAAA,OAAcC;AACpB,QAAMC,WAAAA,MAAAA,IAAmBF,IAAAA;AACzB,QAAA,cAAOE,kBAAAA,QAAAA;AACP,QAAA,mBAAA,YAAA,IAAA,gBAAA;AAEF,SAAO;;AAGL,IAAMC,oBAAiBC,CAAAA,WAAsCC;QAC3D,cAAOC,CAAAA;QACP,gBAAkBC,CAAAA,WAAW,iBAAaC;AAC1C,UAAID,CAAAA,cAAeE,aAAe,IAAA,aAAA,OAAA,MAAA,GAAA;UAChC,CAAA,WAAO,UAAA,IAAA,UAAA,OAAA,MAAA,GAAA;AACT,QAAA,eAAA,eAAA;AACIC,aAAAA;;AAEJ,QAAA,cAAA,KAAA;AACIJ,aAAAA;;AAEJ,QAAA,iBAAA,KAAA;AACKI,aAAAA;;AAEL,SAAA,cAAA,OAAA,cAAA,QAAA,iBAAA,KAAA;AACA,aAAO,aAAA,KAAA,WAAA,GAAA;IACT;AAEA,WAAWC;;aAEPX,SAAYY,QAAK;oBAACD,WAAAA,GAAAA;kBAAM,KAAA;QACxB;MACF,CAAA;AACA;IACA;UACEE,iBAAmB,YAACF,GAAAA,EAAAA;QACpB,cAAA,eAAA,GAAA,EAAA,GAAA,KAAA,GAAA;AACF,qBAAA,KAAA,KAAA;AACAX;;gBAAwB,KAAA;MAC1B;IAEA,CAAA;EACA;AAEF,SAAO;;IAEHQ,mBAAqBA,CAAAA,UAAa;QAClCM,gBAAaC;IACbC,QAAOC,MAAMC,GAAG,CAAA,EAAIF,OAAK,MAAA,GAAA,EAAA,CAAA;IACzBG,OAAMF,MAAMC,OAAOC,CAAAA,KAAI,UAAA,MAAA,MAAA,OAAA,CAAA,IAAA,MAAA;IACvBC,OAAOH,MAAMC,GAAG,CAAA,EAAIE;IACpBC,MAAKJ,MAAMC,GAAI,CAAA,EAAIG;IACrB,OAAA,MAAA,GAAA,CAAA,EAAA;IAEA,KAAK,MAAMV,GAAAA,EAASM,EAAAA;;aAEhBK,SAAAA,MAAkB,MAAIX,CAAAA,GAAAA;AACxB,QAAA,MAAO,KAAA,WAAA,GAAA,GAAA;AACLW,oBAAcH,QAAQ,MAAMR,KAAAA,QAAU,OAAA,EAAA;IACxC,OAAA;AACF,oBAAA,QAAA,MAAA,MAAA;IAEA;EACA;;;",
+  "names": ["LanguageModel", "Effect", "Schema", "AiService", "Obj", "Message", "Schema", "Filter", "Query", "EID", "EntityId", "ReferencedQuotes", "Struct", "references", "Array", "id", "String", "quote", "pipe", "mutable", "annotations", "description", "findReferences", "nouns", "db", "searchKind", "normalizedNouns", "from", "Set", "map", "noun", "trim", "filter", "length", "sort", "leftNoun", "rightNoun", "Promise", "all", "objects", "query", "select", "text", "type", "run", "toString", "flat", "insertReferences", "quotes", "isValid", "left", "right", "regex", "RegExp", "replace", "replaceOutsideLinks", "make", "entityId", "MARKDOWN_LINK", "replacement", "result", "plainStart", "link", "matchAll", "index", "slice", "EXTRACTION_MODEL", "ProperNouns", "Struct", "properNouns", "Array", "String", "annotations", "description", "extractProperNouns", "text", "gen", "value", "scoped", "generateObject", "schema", "prompt", "join", "map", "noun", "trim", "filter", "length", "pipe", "provide", "AiService", "model", "enrichTranscriptMessage", "message", "db", "blocks", "forEach", "block", "_tag", "nouns", "references", "promise", "findReferences", "insertReferences", "Obj", "make", "Message", "id", "sender", "created", "properties", "pipeline", "log", "_ner", "ner", "entities", "text", "tokenGroups", "createTokenGroups", "combinedEntities", "canBeCombined", "prevToken", "currentToken", "currentState", "prevEntity", "entity", "currentEntity", "prevState", "token", "push", "lastTokenGroup", "score", "reduce", "index", "group", "at", "word", "start", "end", "combinedToken"]
 }