brain-mcp

Instructions

    Search across ALL sources: conversations, GitHub, markdown.
    Returns integrated timeline of thinking on a topic.
    

Implementation Reference

• brain_mcp/server/tools_search.py:369-447 (handler)

  The 'unified_search' tool implementation, which queries conversations (vector), GitHub commits (SQL/keyword), and Markdown documentation (SQL/keyword), then aggregates and returns the results.

  def unified_search(query: str, limit: int = 15) -> str:
      """
      Search across ALL sources: conversations, GitHub, markdown.
      Returns integrated timeline of thinking on a topic.
      """
      cfg = get_config()
      results = []
  
      # 1. Conversation embeddings (semantic)
      try:
          embedding = get_embedding(query)
          if embedding and cfg.lance_path.exists():
              lance_results = lance_search(embedding, limit=5)
              for title, content, year, month, sim in lance_results:
                  date = f"{year}-{month:02d}"
                  results.append(("conversation", title or "Untitled", content, date, sim))
      except Exception:
          pass
  
      # 2. GitHub commits (keyword)
      try:
          gh_db = get_github_db()
          if gh_db and cfg.github_commits_parquet.exists():
              gh_results = gh_db.execute("""
                  SELECT 'github' as source,
                         repo_name || ': ' || LEFT(message, 80) as title,
                         message as content,
                         CAST(timestamp AS VARCHAR) as date,
                         0.4 as score
                  FROM github_commits
                  WHERE message ILIKE ? OR repo_name ILIKE ?
                  ORDER BY timestamp DESC
                  LIMIT 3
              """, [f"%{query}%", f"%{query}%"]).fetchall()
              results.extend(gh_results)
      except Exception:
          pass
  
      # 3. Markdown docs (keyword)
      try:
          md_db = get_markdown_db()
          if md_db:
              md_results = md_db.execute("""
                  SELECT 'markdown' as source,
                         COALESCE(title, filename) as title,
                         LEFT(content, 500) as content,
                         CAST(modified_at AS VARCHAR) as date,
                         0.45 as score
                  FROM markdown_docs
                  WHERE content ILIKE ? OR title ILIKE ? OR filename ILIKE ?
                  ORDER BY depth_score DESC
                  LIMIT 3
              """, [f"%{query}%", f"%{query}%", f"%{query}%"]).fetchall()
              results.extend(md_results)
      except Exception:
          pass
  
      if not results:
          return f"No results found across any source for: {query}"
  
      results.sort(key=lambda x: -x[4])
  
      output = [f"## Unified Search: \"{query}\"\n"]
      output.append(f"Found {len(results)} results across sources:\n")
  
      current_source = None
      for source, title, content, date, _ in results[:limit]:
          if source != current_source:
              output.append(f"\n### {source.upper()}")
              current_source = source
  
          date_str = date[:10] if date else "unknown"
          preview = (content[:150] + "...") if content and len(content) > 150 else (content or "")
          output.append(f"**[{date_str}]** {title}")
          if preview:
              output.append(f"> {preview}")
          output.append("")
  
      return "\n".join(output)