Google Genkit Integration
This workspace uses Google Genkit for AI-powered features across multiple modules.
Architecture
an5Agent
Genkit v1.39
RAG Pipeline
-
embedder.tsOpenAI / Cohere / fallback -
indexer.tsSchema + query indexer -
index.tsSingleton vector store
7 Consolidated Tools
-
schema-tools.tslist, describe, relations -
query-tools.tsgenerate, explain, validate -
database-tools.tsexecute, describe, health -
codegen-tools.tsgenerateClientCode, analyze -
rag-tools.tsretrieve schema & queries -
task-tools.tscreate, list, update, delete
RAG Persistent Data
-
__db_an5-schema.jsonSchema vector index -
__db_an5-queries.jsonQuery sample index
an5Tasks
Genkit v1.39
Genkit Flows
-
parseReviewToTasksFlowRegex task extraction -
aiParseReviewToTasksFlowLLM-powered extraction
Task Tools
-
createTaskToolCreate from review issues -
listTasksToolList & filter tasks -
updateTaskToolUpdate status / priority -
deleteTaskToolDelete a task
Setup
Environment Variables
# LLM Provider (for aiParseReviewToTasksFlow)
LLM_PROVIDER=openai # openai | gemini | custom
LLM_API_KEY=sk-...
LLM_MODEL=gpt-4o-mini
LLM_ENDPOINT= # Optional: custom endpoint
Embedding settings for the RAG pipeline are not read from
EMBEDDING_*env vars. They are loaded from theEmbeddingConfigmodel/table viagetEmbeddingConfig()inan5Agent/src/rag/embedder.ts.
Initialize RAG Index
cd an5Agent
npm run rag:index
This indexes .an5 schema files and query-samples.json into the local vector store.
Usage
an5Agent — RAG Retrieval
import { retrieveSchema, retrieveQuerySamples } from '@an5/agent';
// Retrieve schema context for a user question
const schemaContext = await retrieveSchema('How do I find users with their orders?');
// Retrieve similar query samples
const querySamples = await retrieveQuerySamples('aggregate functions');
an5Tasks — Task Extraction
import { createTasksFromReview } from 'an5-tasks';
// Regex-based (fast, no LLM)
const tasks = await createTasksFromReview(reviewText, workspaceDir);
// AI-powered (smarter extraction)
const tasks = await createTasksFromReview(reviewText, workspaceDir, true);
an5Tasks — Task Management
import { getTasks, updateTask } from 'an5-tasks';
// List all todo tasks
const todoTasks = await getTasks(workspaceDir, { status: 'todo' });
// Mark task as done
await updateTask(workspaceDir, 'TASK-1234', { status: 'done' });
Genkit Features Used
| Feature | Module | Purpose |
|---|---|---|
genkit() |
an5Agent, an5Tasks | Initialize Genkit instance |
defineFlow() |
an5Tasks | Define typed task extraction flows |
defineTool() |
an5Tasks | Register task management tools |
runFlow() |
an5Tasks | Execute flows with type safety |
generate() |
an5Tasks | LLM-powered task extraction |
embedder() |
an5Agent | Custom embedding with fallback |
devLocalVectorstore |
an5Agent | Local filesystem vector store |
Document.fromText() |
an5Agent | Create indexed documents |
ai.index() |
an5Agent | Index documents to vector store |
ai.retrieve() |
an5Agent | Semantic retrieval from vector store |
Genkit Features Available (Not Yet Used)
| Feature | Module | Potential Use |
|---|---|---|
definePrompt() |
an5Agent | Reusable prompt templates for code review |
@genkit-ai/firebase |
an5Agent | Production vector store (Firestore) |
@genkit-ai/google-cloud |
an5Agent | GCP telemetry, tracing, logging |
| Model plugins | an5Cli | Replace raw HTTP LLM calls |
| Session/memory | an5Agent | Multi-turn conversations |
| Streaming | an5Cli | Real-time LLM output in Web UI |
| Eval framework | an5Agent | Measure RAG quality |
| Reranking | an5Agent | Improve retrieval quality |
Cross-Repo Integration
an5Cli
LLM Calls
Uses
generate() for:
- Commit message generation
- Changelog generation
- Code review analysis
- Documentation generation
an5Agent
RAG & Tools
Uses Genkit for:
- Schema & query indexing
- Semantic retrieval (RAG)
- Intelligent tool execution
- Natural language routing
an5Tasks
Task Mgmt
Uses Genkit for:
- Task extraction flows
- AI review task parser
- Task management tools
- Bridge tool integration
Vector Store
The local vector store persists to .genkit/vectorstore/ in each module. Index data is also saved as __db_*.json files for portability.
Re-indexing
After schema changes, re-run:
npm run rag:index -w an5Agent
Production Vector Store
For production, replace devLocalVectorstore with:
@genkit-ai/firebase(Firestore)- Pinecone plugin
- Weaviate plugin