Obi - Mise en place for your data

Chat
Chat
Chat Response
Files
File Set-up
Settings
Privacy
Logo
Logo
We tried to gather our thoughts and pitch using FigmaMake to brainstorm and visualize.
Architecture Diagram

From Flicker to Flow

Modern workflows are fragmented.

You jump between folders, notes, and half-remembered files just to find one piece of information. That constant flicker of context switching breaks focus and slows everything down.

We built Obi to fix that.

Inspired by mise en place, the idea of preparing everything before you begin, Obi keeps your data organized and ready so you can stay in a single loop:

ask → retrieve → continue

No digging. No switching. Just flow.

What Obi Does

Obi is a local-first “second brain” for your files.

It instantly finds the most relevant information from your own data, including text and images, without interrupting your workflow.

Instead of searching manually, you just ask. Obi retrieves exactly what you need.

By combining keyword search and semantic search, Obi delivers accurate, context-aware results while reducing friction for users.

It also acts as a context filter for AI agents by providing only the most relevant information instead of entire files. This significantly reduces token usage and improves response quality.

How It Works

Obi runs entirely on your machine and is built around three core layers:

1. Live Indexing

Monitor folders of Markdown, text, and image files
Automatically update as files change
Always reflect real-time data, not stale snapshots

2. Hybrid Retrieval

Keyword search for exact matches
Vector search for semantic understanding
Combined ranking for higher accuracy

Text is chunked, embedded, and stored in a local SQLite database using sqlite-vec.

3. Local AI Inference

Fully local chat and embeddings pipeline
Powered by Gemma via llama.cpp

Models:

Chat: Gemma (quantized GGUF for efficient local inference)
Embeddings: Nomic Embed v2 (GGUF) , CLIP

No cloud. No data leaves your machine.

Why This Matters for Agents

Most agents are inefficient because they operate on too much context.

They either:

Load entire files
Or rely on incomplete keyword search

Obi fixes this by acting as a precision retrieval layer.

Instead of: agent → entire dataset → high token usage

You get: agent → Obi → relevant chunks only

This results in:

Lower token usage
Faster responses
Better grounding and fewer hallucinations
More scalable agent workflows

Obi becomes the mise en place step for agents, preparing exactly the context they need before generating a response.

Retrieval Intuition

We embed both your query and your data into vectors and retrieve the closest matches:

\( \text{cosine-sim}(q, c_i) = \frac{q \cdot c_i}{|q| |c_i|} \)

In simple terms, Obi finds the pieces of your data that are most relevant to your question.

Challenges We Faced

Model startup and compatibility
Loading local models introduced latency and debugging challenges
Speed versus quality tradeoffs
We balanced fast responses with meaningful, grounded results
Hybrid search tuning
Combining keyword and vector search required careful ranking to avoid noise

Accomplishments

Fully local pipeline (hybrid retrieval and inference)
Hybrid search with higher accuracy than standard RAG
Support for both text and image-based context
Real-time indexing with zero manual refresh
A system that reduces both context-switching and token usage

What We Learned

Retrieval matters more than generation
The quality of search has a bigger impact than model size
Efficient context reduces token costs
Smaller, more relevant inputs improve both speed and output quality
Local AI is a systems problem
Performance, memory, and responsiveness are critical
Trust enables flow
Privacy and grounded outputs make users more confident and focused

Built for Local AI with Gemma

Obi is built entirely around on-device intelligence.

Instead of relying on cloud APIs, Obi runs:

Local retrieval
Local embeddings
Local inference using Gemma

This means:

Your data stays private
Your assistant works offline
Your workflow stays uninterrupted

Gemma enables fast, efficient, and context-aware responses directly on a laptop.

Figma Usage

Utilised FigmaMake to iterate on the design language and color schemes of our UI.
We also had a fun time using FigmaMake for making our impactful presentation with playful animations.