Minimal RAG Backend (Text Uploads + GPT-5 Reasoning Token Streaming)

A tiny Go backend that:

Ingests text files, stores them, and generates embeddings (Supabase/Postgres vector column).

Performs RAG: retrieves top-K docs and builds a compact prompt for completion.

Exposes a non-streaming completion endpoint (MVP) and a streaming reasoning-tokens API (WebSocket/SSE) for GPT-5.

Features

Text upload (multipart/form-data) → file persisted locally (MVP) and embedded via OpenAI.

Doc listing for a quick sanity check of what’s retrievable.

RAG completion (non-streaming): builds a short prompt from nearest neighbors and returns a single answer.

Reasoning token streaming (design + endpoint): live token feed via WebSocket/SSE for GPT-5 “thinking” stream.

Strict JSON errors, CORS allow-list, timeouts, small input caps.

API

Base path: /

Health Check

GET /health → {"status":"ok"}

Upload a document

POST /upload (multipart/form-data, field: myFile)

Stores file on disk under ./files/

Creates OpenAI embedding

Inserts row in docs table: { name, location, vector }

Response: {"message":"uploaded","name":"...", "location":"files/upload-..."}

MVP accepts text files for embedding. Non-text will be rejected or you’ll need to extract text upstream.

List docs

GET /docs

Response: {"docs":[{"name":"..."}, ...], "count": N} (MVP counts from results to avoid slow exact count.)

RAG completion

POST /completion Content-Type: application/json

{ "prompt": "Your query here" }

Retrieves top-K (MVP: K=3) by cosine similarity from docs.vector

Builds compact prompt and calls OpenAI chat completion

Response: {"response":"final answer text"}

Reasoning tokens (streaming) — design & contract

Quickstart Prereqs

Go 1.22+

Supabase project (or Postgres with Supabase REST adapter)

OpenAI API key

Env

Create .env:

OPENAI_API_KEY=... POSTGRES_URL=... # Supabase URL POSTGRES_KEY=... # Supabase anon/service key (match your SDK usage)

Database (minimal)

Table: docs

create table if not exists docs ( id bigserial primary key, name text not null, location text not null, vector vector(1536) not null -- Adjust dimension to the model you use (e.g., text-embedding-3-small is 1536). );

Run go run .

Server listens on :8090 by default. Dev CORS allow-list includes http://localhost:5173 and "null" (for file:// testing).

Example requests

Upload:

curl -F "myFile=@notes.txt" http://localhost:8090/upload

List:

curl http://localhost:8090/docs

Non-streaming completion:

curl -X POST http://localhost:8090/completion
-H "Content-Type: application/json"
-d '{"prompt":"What does the design say about streaming tokens?"}'

TODOs

High Priority

• Implement streaming
• Deploy to AWS
• Refactor handlers further into helpers
• Better handling of model output
• Chunking

Medium Priority

• Midleware layer to recover errors, implement global logging, etc.
• Add index to database
• Make frontend look nicer
• Authentication/authorization

Low Priority

• Reformat README
• Unit tests