Unified Enterprise
Intelligence Platform

Powered by Sourcegraph + Amp
Code · Knowledge · Design · Project Management

TECHNICAL PROPOSAL ARCHITECTURE SOW

March 2026  ·  Confidential

16 slides — Swipe ← → or use arrows to navigate

The Challenge

Your requirements span 4 domains — no single tool covers all of them

🎯 Requirements Cross 4 Product Categories

The needs span code intelligence, enterprise knowledge management, design systems, and project management. Today, these exist as isolated silos.

Code

Cross-repo search, semantic navigation, PR impact analysis, code review across 100s of repos

Knowledge

Confluence RAG — 150K pages, 500K attachments, ~100GB of text, images, diagrams, plugins

Design

Figma intelligence — "who uses this component?", design system dependency graph

PM

JIRA Cloud integration — bi-directional sync, agentic task creation from code context

⚠️ Additional Complexity

  • Image/Multimodal RAG — attachments include Visio, Draw.io, screenshots
  • Contradicting sources — 150K pages with overlapping/outdated info
  • Scale — ~100GB of heterogeneous content

Market Reality: No Single Tool Exists

The requirements span 3 distinct product categories that have never been unified

⛔ The Hard Truth

We evaluated every major platform on the market. No single tool in the world addresses 100% of these requirements. The needs cross three fundamentally different product domains that no vendor has ever unified.

Enterprise Search — Competitive Landscape

We evaluated every major enterprise search platform. Here's how they compare:

Capability ✦ Sourcegraph Glean Elastic / OpenSearch Coveo Atlassian Rovo
Cross-repo code search ✅ SCIP semantic ⚠️ Text-only ⚠️ Text-only ❌ Not for code ❌ None
Go-to-def / Find-refs across repos ✅ SCIP native ❌ ❌ ❌ ❌
AI / NLP semantic search ✅ Deep Search 7.0 ✅ Strong ⚠️ ELSER basic ✅ Relevance AI ✅ Rovo AI
Batch Changes across repos ✅ Native ❌ ❌ ❌ ❌
Code review (CI-integrated) ✅ Amp Review ❌ ❌ ❌ ❌
MCP server (agentic access) ✅ GA ❌ ❌ ❌ ❌
Confluence indexing 🔌 MCP extension ✅ Native ✅ Connector ✅ Native ✅ Native
Figma indexing 🔌 MCP extension ⚠️ Basic ❌ ❌ ❌
Image / multimodal search ✅ Deep Search 7.0 ⚠️ Limited ❌ ❌ ⚠️ Basic
Multi code-host support ✅ GitHub, GitLab, BB, Perforce ⚠️ GitHub only ❌ Manual ❌ ⚠️ Bitbucket focus
Scale (repos) ✅ 1M+ repos ⚠️ Limited ✅ Scalable ✅ Scalable ⚠️ Atlassian only

💡 The Verdict

Sourcegraph is the only platform that combines deep code intelligence with extensible MCP integration. Competitors like Glean index Confluence natively but have shallow code understanding. Elastic is powerful for text but has zero code intelligence. Rovo is locked to the Atlassian ecosystem. Only Sourcegraph provides SCIP navigation, Batch Changes, and an MCP server — making it the natural hub to extend into Confluence, Figma, and JIRA via purpose-built integrations.

Solution Architecture

Sourcegraph as the core, extended via MCP Gateway

🖥️ Unified Developer Portal
Web UI · CLI · IDE · Mobile
▼
🔌 MCP Gateway
Router · Auth · Cache · Audit
▼     ▼     ▼     ▼     ▼
📦 Sourcegraph
Code Search · Deep Search · SCIP · Batch Changes · Insights
⚡ Amp Agent
Code Review · Agentic Coding · Checks
📚 Confluence RAG
Crawler · Embeddings · Vector DB · Hybrid Search
🖼️ Multimodal
OCR · Vision LLM · CLIP · Image Search
🎨 Figma Intel
Component Graph · Design Search
▲
📋 JIRA MCP
Bi-directional Sync · Context Enrichment

🔑 Key Design Decisions

  • MCP-native: Every layer exposes tools via MCP — agents query all sources uniformly
  • Sourcegraph at center: Code intelligence is the primary layer; others extend it
  • Incremental: Each layer delivers value independently — no big-bang deployment
  • Agent-agnostic: Works with Amp, Cody, or any MCP-compatible agent

Layer 1: Code Intelligence

Sourcegraph 7.0 + Amp — Production-ready, zero custom development

📦 Sourcegraph Enterprise — Already Built

ComponentStatusCapability
Code Search✅ ProductionRegex, structural, multi-repo, multi-host (GitHub, GitLab, Bitbucket, Perforce)
Deep Search (NLP)✅ ProductionNatural language code queries + image support (v7.0)
SCIP Cross-Repo Nav✅ ProductionGo-to-definition, find-references across ALL repos with version awareness
Batch Changes✅ ProductionApply code changes across 100s of repos in a single operation
Code Insights✅ ProductionMigration tracking, version adoption, custom metric dashboards
Code Monitors✅ ProductionAlert on code patterns appearing in any repo
MCP Server✅ GAAgents query code intelligence via standard MCP protocol

⚡ Amp — Agentic Code Review & Development

FeatureDescription
Amp ReviewCI-integrated code review, posts to PRs, custom Checks per codebase area
Cross-Repo ContextLibrarian sub-agent reads across private repos via GitHub/Bitbucket
OracleDeep reasoning engine for architecture review and complex debugging
Thread SharingTeam-visible conversation threads for knowledge sharing
Skills & MCPExtensible via custom skills and MCP servers

✅ This entire layer is production-ready. Zero custom development needed.

Layer 2: Confluence Knowledge RAG

Custom pipeline for 150K pages + 500K attachments

📚 Data Flow Pipeline

Confluence API
→
Crawler
→
Doc Processor
→
Chunker
→
Embeddings
→
Vector DB
Attachments
→
Tika / Unstructured
→
OCR + Vision
→
CLIP Embed
→
Image Vector DB
Agent Query
→
MCP Gateway
→
Hybrid RAG
→
Re-rank
→
Answer + Sources
ComponentTechnologyPurpose
Confluence CrawlerAtlassian REST API + WebhooksIncremental crawl of all spaces, pages, attachments
Document ProcessorUnstructured.io OSSParse HTML, PDF, DOCX, PPTX → structured text
Chunking EngineLangChain OSSSemantic chunking with metadata preservation
DeduplicationMinHash + SimHashHandle contradicting/duplicate sources
Freshness RankerCustom scoringRank by last-modified, activity, endorsements
Embedding ServiceOpenAI text-embedding-3-largeDense vector embeddings (3072 dims)
Vector DatabaseWeaviate OSSStore ~5M+ chunks with metadata filtering
BM25 IndexElasticsearch OSSSparse keyword search for hybrid retrieval
Hybrid RAG EngineLlamaIndex OSSDense + sparse retrieval with re-ranking
MCP ServerTypeScript MCP SDKExpose RAG as MCP tools for Sourcegraph/Amp

Layers 3–5: Image · Figma · JIRA

Extending Sourcegraph's reach beyond code

🖼️ Layer 3: Multimodal / Image RAG

Images from Confluence
→
Tesseract OCR
→
Vision LLM Caption
→
OpenCLIP Embed
→
Image Vector Store

Handles PNG, SVG, Visio, Draw.io diagrams. Vision LLM generates searchable text descriptions. CLIP enables "find diagrams similar to X" queries.

🎨 Layer 4: Figma Design Intelligence

Figma REST API
→
Component Crawler
→
Dependency Graph (Neo4j)
→
Figma MCP Server

Builds a full component dependency graph. Answer queries like "who uses this component?", "find pages using color X", "which designs reference this token?"

📋 Layer 5: JIRA Integration

JIRA Cloud API
→
JIRA MCP Server
→
Bi-directional Sync
→
Context Enrichment

JQL search, create/update issues via MCP. Attach Sourcegraph code links to JIRA tickets automatically. Amp can create JIRA epics from batch migration results.

Final System — Block Diagram

Complete integrated architecture

CLIENTS
👤 Developer
IDE / CLI / Amp
🤖 CI/CD Pipeline
Automated Agents
🌐 Web Portal
Dashboard
▼   MCP Protocol   ▼
ORCHESTRATION
🔌 MCP Gateway — Router · OAuth2/OIDC Auth · Redis Cache · Query Planner · Audit Logger
▼    ▼    ▼    ▼    ▼
INTELLIGENCE LAYERS
Sourcegraph 7.0
Code Search · Deep Search
SCIP Nav · Batch Changes
Code Insights · Monitors
📚 Confluence RAG
Crawler · Embeddings
Vector DB · BM25
Hybrid Retrieval · Re-rank
🖼️ Multimodal Engine
OCR · Vision Captioning
CLIP Embeddings
Image Vector Search
⚡ Amp Agent
Code Review · Checks
Oracle · Librarian
Thread Sharing
🎨 Figma Intelligence
Component Graph
Design Search
Usage Analytics
📋 JIRA Integration
JQL via MCP
Bi-directional Sync
Code Link Enrichment
▼    ▼    ▼
DATA STORES
🗄️ Sourcegraph DB
SCIP indexes · Repo metadata
🔢 Weaviate / Qdrant
Text + Image vectors
🔍 Elasticsearch
BM25 keyword index
🕸️ Neo4j
Figma component graph
⚡ Redis
Cache + sessions

Agentic Workflow Examples

Real-world scenarios powered by the unified platform

🔍 Workflow 1: "Will this PR break anything?"

Developer opens PR
→
Sourcegraph SCIP: Find all references
→
Confluence: Architecture docs
→
JIRA: Related issues
→
Impact Report

📚 Workflow 2: "How does our payment system work?"

Agent Query
→
Confluence: Design docs
→
Deep Search: Payment code
→
Image RAG: Architecture diagrams
→
Full Explanation

🎨 Workflow 3: "Who uses the Button component?"

Agent Query
→
Figma: 47 design files
→
Sourcegraph: 12 repos import <Button>
→
Unified Usage Report

🔄 Workflow 4: "Migrate lodash across all repos"

Sourcegraph: Find vulnerable versions
→
Confluence: Migration guide
→
Batch Changes: PRs across repos
→
JIRA: Auto-create epic + tasks

Statement of Work (SOW)

Scope, deliverables, assumptions

📄 Project Scope

Design, develop, and deploy a Unified Enterprise Intelligence Platform with Sourcegraph + Amp at its core, extended with Confluence knowledge RAG, Figma design intelligence, JIRA project management, and multimodal image understanding — all unified through an MCP Gateway.

📦 Deliverables

#DeliverablePhase
D1MCP Gateway with routing, auth, caching, audit loggingPhase 1
D2Sourcegraph Enterprise deployment + MCP integrationPhase 1
D3JIRA MCP Server with bi-directional syncPhase 1
D4Confluence crawler + document processor (text only)Phase 2
D5Vector DB + hybrid RAG engine + MCP serverPhase 2
D6Deduplication & freshness ranking systemPhase 2
D7Image extraction + OCR + vision captioning pipelinePhase 3
D8CLIP image embeddings + image vector searchPhase 3
D9Figma component graph + MCP serverPhase 4
D10Unified dashboard / developer portalPhase 5
D11Documentation, runbooks, training materialsPhase 5
D12Load testing report (150K pages, 500K attachments validated)Phase 5

🔒 Assumptions

Project Timeline

5 phases over ~20 weeks

Phase 1 — Foundation (Weeks 1–4)
  • MCP Gateway — Router, auth, caching, audit
  • Sourcegraph deployment — Enterprise setup, repo indexing, MCP server config
  • JIRA MCP — Configure + bi-directional sync
  • ✅ Milestone: Agent can search code + query JIRA through unified MCP
Phase 2 — Confluence RAG (Weeks 5–12)
  • Crawler — Incremental Confluence sync, attachment download
  • Document processing — Unstructured.io pipeline, chunking, metadata
  • Vector DB + RAG — Weaviate setup, hybrid retrieval, re-ranking
  • Dedup + freshness — MinHash dedup, recency scoring
  • ✅ Milestone: Agent can answer questions from 150K Confluence pages
Phase 3 — Multimodal (Weeks 10–15)
  • Image extraction — From Confluence attachments, inline images
  • OCR + Vision — Tesseract + Vision LLM captioning
  • CLIP embeddings — Image vectorization + visual search
  • ✅ Milestone: Agent can search and understand images/diagrams
Phase 4 — Figma (Weeks 13–17)
  • Figma crawler — REST API sync of team files, components, styles
  • Component graph — Dependency graph in Neo4j
  • ✅ Milestone: Agent can query Figma design system dependencies
Phase 5 — Integration & Hardening (Weeks 17–20)
  • Unified dashboard — Cross-layer search portal
  • Load testing — Full 100GB corpus validation
  • ✅ Milestone: Production-ready, all layers connected

Effort & Investment

Team composition and effort breakdown

👥 Team Composition

RoleCountDuration
Tech Lead / Architect120 weeks (full)
Backend Engineers (Python/Go/TS)316–20 weeks
ML / RAG Engineer112 weeks (Phase 2–3)
DevOps / Infrastructure110 weeks
QA Engineer18 weeks (Phase 3–5)

📊 Effort Breakdown by Phase

PhaseDurationPerson-Weeks% of Total
Foundation4 weeks14 pw16%
Confluence RAG8 weeks32 pw38%
Multimodal6 weeks16 pw19%
Figma5 weeks12 pw14%
Integration4 weeks10 pw12%
TOTAL~20 weeks~84 pw100%

💰 Development Investment

TBD

Depends on final assessment of Confluence content complexity, infrastructure requirements, and team composition.

🏗️ Annual Infrastructure

TBD

Depends on final assessment of hosting model (cloud vs. on-prem), LLM API usage volume, and scale requirements.

Risks & Mitigations

Key risks with mitigation strategies

RiskImpactLikelihoodMitigation
Confluence content too heterogeneous (plugins, macros) HIGH MED Early content audit in Phase 2. Build plugin-specific parsers for top 5 most used.
100GB embedding generation takes too long MED MED Batch processing with parallelized workers. Incremental re-indexing after initial load.
Contradicting sources degrade RAG quality HIGH HIGH Freshness ranking + source authority scoring. Show provenance. Allow marking pages as canonical.
Figma API rate limits at scale LOW MED Webhook-based incremental sync. Aggressive caching. Full crawl during off-hours.
Image RAG accuracy insufficient MED MED Multi-signal retrieval: OCR text + vision captions + CLIP vectors combined.
Confluence on-prem connectivity / firewall MED MED Deploy crawler inside network perimeter. Push processed chunks to cloud vector DB.

Success Criteria & KPIs

Measurable outcomes per phase

KPITargetMeasurement
Cross-repo code search latencyp95 < 2 secSourcegraph performance dashboard
Confluence RAG query latencyp95 < 3 secRAG pipeline metrics
Confluence coverage> 95% of 150K pages indexedCrawler completion report
Attachment processing> 90% of 500K attachments parsedDocument processor stats
RAG answer relevance> 80% user satisfaction (top-5)User feedback sampling
Image search precision> 70% relevant in top-3Manual evaluation set
Figma component coverage100% of published componentsFigma API vs. graph DB count
JIRA sync latency< 30 sec webhook-to-updateWebhook processing logs
MCP Gateway uptime> 99.5%Health check monitoring
End-to-end cross-layer query< 8 sec totalOrchestrator tracing

📈 Expected Business Impact

Before:
  • 30+ min/day searching for code context
  • Confluence knowledge siloed, hard to find
  • PR impact analysis = manual + faith-based
  • Design-code alignment = tribal knowledge
After (with Sourcegraph):
  • < 5 min — agent finds code + docs + design in one query
  • All 150K pages searchable semantically
  • Automated cross-repo impact reports on every PR
  • "Who uses this component?" — seconds (code + Figma)

Why Sourcegraph Is the Right Foundation

The only platform that can anchor this solution

🔍 Unmatched Code Intelligence

  • SCIP-powered cross-repo navigation — no other tool has this
  • Deep Search with NLP + image support (v7.0)
  • Supports GitHub, GitLab, Bitbucket, Perforce simultaneously
  • Scales to 1M+ repositories

🔄 Batch Changes

  • Apply code changes across 100s of repos in one operation
  • Security patches, dependency upgrades, API migrations
  • No other platform offers this at enterprise scale

🔌 MCP-Native

  • GA MCP server — the natural orchestration hub
  • Any AI agent can query Sourcegraph intelligence
  • Future-proof: works with Amp, Claude, GPT, Cursor

⚡ Amp Agent

  • CI-integrated code review with custom Checks
  • Oracle for deep reasoning, Librarian for cross-repo research
  • Thread sharing for team knowledge
  • Skills + MCP extensibility

🏢 Enterprise Trust

  • Used by Databricks, Uber, Lyft, Reddit, GE, Dropbox, Palo Alto Networks, Leidos
  • SOC 2 Type II · Zero data retention on LLM inference · SSO (Okta, SAML)
  • Self-hosted, cloud, or air-gapped deployment options
  • Backed by Andreessen Horowitz, Sequoia Capital

Next Steps

From proposal to production

1️⃣ Technical Discovery (1 week)

Confluence content audit, Figma team inventory, infrastructure review, final assessment

2️⃣ Architecture Validation (1 week)

PoC with 1,000 Confluence pages → validate RAG quality before full scale

3️⃣ Scoping & Kickoff (1 week)

Finalize SOW based on discovery findings, Sourcegraph licensing, team onboarding

4️⃣ Phase 1 Start

MCP Gateway + Sourcegraph + JIRA — first value delivered in 4 weeks

Let's build the intelligence layer your engineering team deserves.

Contact: amp-devs@ampcode.com  |  sourcegraph.com/contact

Confidential — March 2026

1 / 16
← swipe or use arrow keys →