Tara

Status: ✅ Production Type: Engineering assistant · Self-improving AI agent Stack: Slack Bolt + NeuroLink SDK + Vertex AI + Redis + PostgreSQL

Purpose

Tara is a Slack-native AI assistant for engineering teams. She receives messages and file attachments in Slack, routes them through NeuroLink's pipe with full MCP tool access (Bitbucket, JIRA, GitHub, Figma, OpenObserve), and responds with streaming answers or structured PDF reports. Each Slack thread gets its own isolated NeuroLink instance with Redis-backed conversation memory.

Stream Types

Stream	Direction	Description
Slack DM messages	→ Pipe	User text + file attachments via Slack Assistant API
@mention events	→ Pipe	Channel messages where user mentions @tara
Attached files	→ Pipe	PDFs, images, code files, CSV, Word — 16+ types via ProcessorRegistry
Streaming tokens	Pipe →	Real-time token stream posted to Slack thread
PDF reports	Pipe →	Structured JSON → PDF, uploaded to Slack
Tool results	Pipe → Slack	PR data, JIRA tickets, code search, Figma assets

Input / Output Contract

Input (Slack Events)

{
  // User message
  text: string,
  user: string,          // Slack user ID
  channel: string,       // Channel or DM ID
  ts: string,            // Thread timestamp
  thread_ts?: string,    // Parent thread if reply

  // Optional file attachments
  files?: Array<{
    name: string,
    mimetype: string,    // Auto-detected by FileDetector
    url_private: string  // Slack download URL
  }>
}

Output (Slack messages)

// Streaming response — posted progressively
{
  text: string,          // Accumulated tokens
  blocks: Block[],       // Slack block kit UI
  thread_ts: string      // Always threaded
}

// PDF report (for complex analysis)
{
  filename: "tara-report-{timestamp}.pdf",
  content: Buffer,       // PDF binary
  initial_comment: string  // Summary message
}

NeuroLink Integration

import { NeuroLink } from "@juspay/neurolink";

// Per-thread instance — LRU pool, max 100 concurrent
const neurolink = new NeuroLink({
  defaultProvider: "vertex",
  memory: {
    type: "redis",
    prefix: `thread:${sessionId}:`,
  },
  observability: {
    langfuse: { enabled: true },
  },
});

// Streaming response
for await (const token of neurolink.stream({
  prompt: userMessage,
  files: attachedFiles, // Processed by FileDetector + ProcessorRegistry
  tools: { ...mcpTools }, // Bitbucket, JIRA, GitHub, Figma, OpenObserve
  memory: { enabled: true },
})) {
  await slackClient.chat.update({ text: accumulated });
}

Features used:

• neurolink.stream() — streaming Slack responses
• neurolink.generate() — structured JSON for PDF reports
• Redis conversation memory (per-thread isolation)
• Multimodal file processing (FileDetector + ProcessorRegistry — 16+ file types)
• MCP servers: Bitbucket Server, JIRA, GitHub, Figma, OpenObserve
• Langfuse observability with session/user/conversation context enrichment

MCP Tools

Server	Tools Available
Bitbucket Server	PR review, branch ops, code search, file content, diffs
JIRA	Ticket ops, issue search, project queries
GitHub	Repository data
Figma	Design files, component inspection
OpenObserve	Traces, logs, metrics queries

Gateway Unlocked

Tara gives engineering teams:

• Codebase Q&A — ask questions, get answers with code references
• PR assistance — review PRs, generate descriptions, search context
• JIRA integration — create, update, query tickets from Slack
• Autonomous tasks — multi-step coding tasks via tool loops
• PDF reports — long-form analysis delivered as downloadable documents

Operational Notes

• Latency: 6.9s–70.3s — exceeds Slack's 3s ACK deadline. Handled via unhandledRequestHandler returning 200 immediately, then posting response when ready
• Concurrency: LRU pool of 100 NeuroLink instances, one per active Slack thread
• File processing: All files downloaded and processed locally before sending to AI — unsupported formats return helpful error messages
• Async tasks: BullMQ task queue for long-running operations (title generation, DB sync)
• Session persistence: Thread titles auto-generated and stored in PostgreSQL; conversation state in Redis
• Observability: Full OpenTelemetry traces, Langfuse AI spans with userId + sessionId + conversationId context