Second
Brain

17  prepareText(text: string): string {18    if (text.length <= this.maxChars) return text;19    // Truncate at paragraph boundary if possible to preserve semantic coherence20    const truncated = text.slice(0, this.maxChars);21    const lastPara = truncated.lastIndexOf('\n\n');22    return lastPara > this.maxChars * 0.7 ? truncated.slice(0, lastPara) : truncated;23  }24 25  async embed(text: string): Promise<Float32Array> {26    const prepared = this.prepareText(text);27    const res = await fetch(`${this.host}/api/embeddings`, {28      method: 'POST',29      headers: { 'Content-Type': 'application/json' },30      body: JSON.stringify({ model: this.model, prompt: prepared }),31    });32    if (!res.ok) {33      throw new Error(`Ollama embedding failed: ${res.status} ${res.statusText}`);34    }35    const data = await res.json() as { embedding: number[] };36    return new Float32Array(data.embedding);37  }

3export function cosineSimilarity(a: Float32Array, b: Float32Array): number {4  let dot = 0, normA = 0, normB = 0;5  for (let i = 0; i < a.length; i++) {6    dot += a[i] * b[i];7    normA += a[i] * a[i];8    normB += b[i] * b[i];9  }10  const denom = Math.sqrt(normA) * Math.sqrt(normB);11  return denom === 0 ? 0 : dot / denom;12}

72    const limit = rawLimit ?? 10;73    let results;74    if (ollamaAvailable) {75      const queryEmbedding = await ollama.embed(query);76      results = findTopK(queryEmbedding, embeddingsCache, limit, { project, type });77    } else {78      // Keyword fallback79      const notes = listNotes(VAULT_PATH).map(p => {80        const n = readNote(VAULT_PATH, p);81        return { path: p, title: n.title, tags: n.frontmatter.tags ?? [], content: n.content };82      });83      results = keywordSearch(query, notes).slice(0, limit);84    }