RAG Wrapper Patterns

Patterns for augmenting any agent with Qdrant context retrieval.

Quick Start

To wrap an agent with RAG:

Use rag-proxy agent:
  Target: {agent-to-wrap}
  Task: {the task}

Manual Wrapping Pattern

If you need custom control, follow this pattern:

Step 1: Query Relevant Context

Tool: qdrant-find
Query: {key terms from task}

Step 2: Format Context Block

## Retrieved Context

### Source: {metadata.source}
Harvested: {metadata.harvested_at}
Type: {metadata.type}

{document content}

---

Step 3: Prepend to Task

{context blocks}

## Task

{original task}

---
Note: Above context is from stored knowledge. Verify if needed.

Step 4: Delegate

Tool: Task
Agent: {target-agent}
Prompt: {enriched prompt}

Enriched Prompt Template

# Context from Stored Knowledge

The following relevant information was retrieved from project memory:

{{#each contexts}}
## From Qdrant
**Source:** {{metadata.source}}
**Harvested:** {{metadata.harvested_at}}

{{content}}

---
{{/each}}

# Your Task

{{original_task}}

---

**Note:** The context above comes from previously harvested research.
Use it if relevant, but verify currency for time-sensitive information.
The `harvested_at` dates indicate when the content was stored.

Selective Wrapping

Not all tasks need RAG. Skip for:

| Task Type | Wrap? | Reason | |-----------|-------|--------| | Fresh research | No | Need current, not cached data | | Simple edits | No | Context not needed | | RAG-aware agents | No | Already query Qdrant | | Implementation | Yes | Benefit from patterns, decisions | | Debugging | Yes | Previous solutions may help | | Architecture | Yes | Decisions and constraints matter |

Agent-Collection Affinity

Map agent types to useful query topics:

| Agent Type | Query Topics | |------------|--------------| | frontend-developer | react, design system, components | | backend-architect | api, architecture, decisions | | security-auditor | security, authentication, vulnerabilities | | devops | infrastructure, terraform, deployment | | tester | testing, coverage, quality |

Storing Results

After the target agent completes:

Tool: qdrant-store
Information: "<valuable findings>"
Metadata:
  source: "agent-output"
  type: "generated"
  harvested_at: "<ISO date>"
  tags: "<relevant,keywords>"

Error Handling

| Scenario | Action | |----------|--------| | Empty query results | Proceed without context | | Qdrant unavailable | Fall back to unwrapped delegation | | Target agent fails | Report error, don't retry with less context |