Features

Everything you need to
build better RAG systems

A complete toolkit for designing, executing, debugging, and collaborating on retrieval-augmented generation pipelines.

Visual Canvas

Design your pipeline,
node by node

Drag nodes from the palette, connect them with typed wires, and configure everything inline. Your entire RAG architecture in one visual workspace.

8 specialized node types

Query, HyDE, Embedding, Retriever, Reranker, Judge, LLM, and Agent — each purpose-built.

Typed connections

STRING, VECTOR, DOCUMENT, JSON — the canvas validates wire compatibility as you build.

Inline property panels

Click any node to configure model, temperature, top-k, and prompt right on the canvas.

Keyboard shortcuts

Cmd+C/V to copy nodes, Delete to remove, Cmd+Z to undo. Feels like a native app.

crag-pipeline.hachi
saved
Query
STRING
HyDE
VECTOR
Embedding
Retriever
DOCUMENT
Reranker
JSON
LLM
6 nodes · 5 connections
CRAG pattern
valid

Smart auto-layout

Nodes automatically arrange into clean topological layouts. Manual adjustments snap to grid.

Export to code

One-click export to a standalone TypeScript pipeline you can run anywhere.

Version history

Every save creates a version. Diff, compare, and restore any previous pipeline state.

Template library

Start from pre-built patterns: HyDE, CRAG, Fusion, Parent-Child, and more.

query → embedding → retriever → judge → reranker → llm → output

wire-tap-debug.hachi
inspecting
Query
Embed
Retriever
5 docs · 0.94 · 156ms
Reranker
inspecting: retriever → reranker5 documents captured
Retriever Outputapplication/json
{
  "documents": [
    { "title": "RAG Overview", "score": 0.94, "tokens": 312 },
    { "title": "Vector Search", "score": 0.91, "tokens": 287 },
    { "title": "Embedding Models", "score": 0.87, "tokens": 245 }
  ]
}
Wire Tap

Click any wire.
See everything.

Every connection is inspectable. Click a wire to see exactly what data flowed through it — documents, scores, embeddings, latency, and cost. No logging setup required.

Inspect any connection

Click any wire to see the full payload. Documents, vectors, scores — everything in structured JSON.

Embedding vector preview

Visualize high-dimensional vectors as sparklines. Spot drift and quality issues instantly.

Cost tracking per step

Every node reports token usage and estimated cost. See exactly where your budget goes.

Latency waterfall

Timeline view of every step. Identify bottlenecks in your pipeline at a glance.

Real-time streaming

Watch data flow through your pipeline live as each node completes. No refresh needed.

Run comparison

Compare outputs between runs side-by-side. Track improvements across iterations.

Export results

Export Wire Tap traces as JSON for automated testing and regression suites.

retriever → 5 docs · score: [0.94, 0.91, 0.87, 0.84, 0.81] · 450ms · $0.0002

Execution

Not a simulation.
Real API calls.

When you press Run, hachi calls your actual LLM and vector store APIs with your credentials. Real tokens, real latency, real results — not a sandbox mockup.

Your models, your keys

Bring your own OpenAI, Anthropic, or Cohere API keys. Your credentials, your data.

Parallel execution

Independent branches run concurrently. Topological sort ensures correct dependency order.

Streaming via SSE

Results stream back in real-time via server-sent events. Watch nodes complete live.

Automatic retries

Rate limits and transient failures are retried with exponential backoff. No lost runs.

crag-pipeline.hachi
executing
"How does retrieval-augmented generation work?"
Query
HyDE
Embedding
Retriever
Reranker
Judge
LLM
elapsed: 1.85stokens: 2,847
est. cost: $0.0142

Topological execution

Nodes execute in dependency order. Parallel branches run concurrently for maximum throughput.

Multi-provider

Mix OpenAI, Anthropic, and Cohere in the same pipeline. Each node uses its own provider.

Cost tracking

Real-time token counting and cost estimation. Per-node and total pipeline cost breakdown.

Retry & fallback

Automatic retries with exponential backoff. Configure fallback models for resilience.

query:200ms → embed:350ms → retrieve:450ms → rerank:600ms → judge:800ms → llm:1.5s = 3.9s total

team-pipeline.hachi
A
B
C
3 online
Query
Embed
Retriever
Reranker
LLM
5 nodes · all passing
synced
Activity
12:04:31
A
Alice updated LLM temperature to 0.8
12:04:28
B
Bob added Reranker node
12:04:25
C
Carol opened Wire Tap on Retriever
Collaboration

Build together,
in real time.

See your teammates' cursors on the canvas, watch them configure nodes, and share Wire Tap results — all in real time with zero setup.

Live cursors with presence

See who's on the canvas and what they're working on. Colored cursors with name labels.

Conflict-free sync (Yjs CRDT)

Concurrent edits merge automatically. No locking, no conflicts, no lost work.

Shared Wire Tap

When one person inspects a wire, everyone can see the results. Debug together.

Org-scoped access control

Owner, Admin, Editor, Viewer — four granular roles with fine-grained permissions.

Team awareness

Live presence indicators show who's online. See which node each teammate is editing.

Role-based access

Four granular roles control who can edit, execute, and manage pipeline settings.

Activity feed

Timestamped log of every change. See who did what and when across the team.

yjs crdt · webrtc p2p · <50ms sync · conflict-free merging

RAG Patterns

Advanced patterns, pre-built

Pre-built nodes for sophisticated retrieval patterns. Each encapsulates a proven technique with real performance benchmarks.

+15-30% recall

HyDE

Hypothetical Document Embeddings

Generate a hypothetical answer to improve embedding quality for short queries.

  • Better retrieval for short queries
  • Reduces embedding mismatch
  • Works with any embedding model
Try in mini map →
2x context

Parent-Child

Hierarchical Chunking

Match on small, precise chunks but return their parent chunks for context.

  • Precise matching
  • Rich context
  • Reduces hallucination
Try in mini map →
+12% MRR

Fusion

Reciprocal Rank Fusion

Combine multiple retrieval methods using reciprocal rank fusion.

  • Hybrid search
  • Better recall
  • More robust results
Try in mini map →
-40% hallucination

Judge (CRAG)

Corrective RAG

Evaluate retrieved documents for relevance before passing to the LLM.

  • Self-correcting
  • Quality assurance
  • Fallback handling
Try in mini map →
3x precision

Reranker

Cross-encoder Reranking

Use a cross-encoder model to rerank initial retrieval results.

  • Higher precision
  • Better top-k results
  • Cross-encoder accuracy
Try in mini map →
+50% accuracy

Agentic RAG

Agent with Retrieval Tool

Give an LLM agent access to retrieval as a tool for dynamic, multi-step search.

  • Dynamic retrieval
  • Multi-step reasoning
  • Self-directed search
Try in mini map →
Integrations

Works with your stack

Connect to the models and vector stores you already use. OpenAI API compatible.

LLMs & Embeddings

OpenAInative

GPT-4o, o1, text-embedding-3

Anthropicnative

Claude 3.5 Sonnet, Claude 3 Opus

Coherenative

Command R+, Rerank v3, embed-v3

Ollamacompatible

Llama 3.1, Mistral, self-hosted

Vector Stores

Pineconenative

Serverless, namespaces, metadata

Weaviatenative

Hybrid search, multi-tenancy

Qdrantnative

Filtering, sparse vectors

PostgreSQLnative

pgvector, HNSW, IVFFlat

Plus Chroma, Milvus, Azure OpenAI, AWS Bedrock, and any OpenAI-compatible API.