Cheeserag Studio | Pomai Ecosystem

What's Inside

Cheesebrain

C++20 | LLM Inference

OpenAI-compatible server (/v1/chat/completions, /v1/embeddings).

PomaiDB

C++20 + Python | Vector DB

Multi-membrane edge vector DB with zero-OOM guarantees.

Cheese API

Python / FastAPI | Orchestrator

Workspace CRUD, async ingest, citation metadata, audio overviews.

Cheesepath Agent

Go | Autonomous Agent

CLI agent with ReAct, planning, multi-role panel, tool registry.

Studio UI

TypeScript / Next.js 14 | UI

3-panel workspace: sources, chat, notes with exact PDF citations.

Edge AI Design Philosophy

"I could have easily integrated the OpenAI API, but my goal was to engineer a highly secure, air-gapped, local-first RAG workspace capable of running on resource-constrained hardware. By designing a micro-agent pipeline architecture and integrating it tightly with PomaiDB, I successfully mitigated the reasoning limitations of a 0.5B model. This kept the total memory footprint under 1 GB while maintaining high extraction accuracy and zero data leakage."

Tactic 1: Algorithmic Citation

The LLM is never asked to place citation markers. The backend runs TF-IDF cosine similarity between the generated answer and retrieved chunks. Footnote markers are programmatically inserted, ensuring zero hallucinated citations.

Tactic 2: Prompt Chaining

Audio overviews run sequentially: extract bullet points, aggregate via pure Python, and synthesize dialogue. Each LLM call is kept under 512 tokens, well within the reliable context window of a 0.5B model.

Tactic 3: Constrained Generation

Enforcing max_tokens to 150 prevents rambling. Using completion-style prompts ending with a colon forces the model to fill a blank rather than drift into uncontrolled generation.

Recommended Start (Docker Compose)

The easiest path. Docker builds all three C++ submodules automatically inside containers.

docker-compose up --build -d

Service	URL
Studio Web UI	http://localhost:3000
Cheese API	http://localhost:9090/docs
Cheesebrain	http://localhost:8080

Manual Build & Installation

1. Submodules & PomaiDB

git clone https://github.com/pomagrenate/cheeserag.git cd cheeserag git submodule update --init --recursive cd third_party/pomaidb git submodule update --init third_party/palloc cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=g++ -DPOMAI_BUILD_TESTS=OFF cmake --build build -j$(nproc) cd ../..

2. Cheesebrain & Go Agent

cd third_party/cheesebrain cmake -B build -DCMAKE_BUILD_TYPE=Release cmake --build build --config Release -j$(nproc) cd ../.. go build -o build/cheeserag-agent ./cmd/cheeserag-agent/

3. Python API Setup

python3 -m venv .venv source .venv/bin/activate pip install --upgrade pip pip install -r requirements.txt export POMAI_C_LIB=$(pwd)/third_party/pomaidb/build/libpomai_c.so export PYTHONPATH=$(pwd)/third_party/pomaidb/python:$PYTHONPATH

Minimum Prerequisites

CMake: 3.20+
Compiler: C++20 capable (GCC 11+, Clang 14+)
Go: 1.23+
Python: 3.10+
Node.js: 18+

System Data Flow

Browser
  │  drag-drop PDF
  ▼
Studio (Next.js :3000)
  │  POST /api/v1/ingest  (multipart)
  ▼
Cheese API (FastAPI :9090)
  │  1. process_file_with_meta()
  │  2. fetch_embedding()
  │  3. put_chunk_with_text()
  │  4. store_chunk_meta()
  │  SSE progress → browser
  ▼
PomaiDB (libpomai_c.so — in-process)

                              Chat query
Browser ──────────────────────────────────────►
                                                Cheese API
                                                  │  embed query
                                                  │  search_rag_membrane()
                                                  │  if max_score < 0.35 → "not found"
                                                  │  else: build grounded system prompt
                                                  │  stream /v1/chat/completions
                                                  ▼
                                               Cheesebrain (:8080)