nadir

Semantic document search engine. Ingests text files, chunks + embeds them locally, stores in Qdrant, serves hybrid semantic+keyword search over HTTP, with optional cross-encoder reranking and LLM answer generation.

A 60-second recording at 2× speed (click the picture to play): delete all chats, import two Markdown files, ask a question, open Inspect and Inspect context sent to the LLM, click a citation, ask a follow-up, then ask a question the documents cannot answer.

Run it (Linux or macOS)

Install the tools below, then run these steps from the repository root. The two terminals stay open.

# 1. Models (Ollama must be installed and running). These are the defaults in
#    config.yaml and can be changed, see "Choose your models".
ollama pull embeddinggemma-300m-q8
ollama pull qwen3.5:4b

# 2. Terminal A: Qdrant + API on http://localhost:8100
./scripts/local.sh

# 3. Terminal B: dashboard on http://localhost:3002
cd web/dashboard && npm ci && npm run dev

Open http://localhost:3002, click + → Import Markdown or PDF, choose a .md file (try eval/samples/cache.md), then ask a question about it.

	Linux	macOS
Podman	./scripts/setup_podman_host.sh (Ubuntu 24.04+), or your package manager	brew install podman podman-compose, then podman machine init && podman machine start
Go 1.27+, Python 3.12+, Node 24 LTS	Package manager or the official installers	brew install go python@3.12 node@24
Ollama	ollama.com	ollama.com

The macOS column and local.sh on a Mac are untested here; Podman runs inside a podman machine VM there, see the Compose guide.

Choose your models

The two ollama pull commands pull the defaults, not requirements. Each model is set in config.yaml; pull whichever model you choose and put its name in the matching key:

Role	Config key	Default
Embeddings	embedder.model (and embedder.dimensions)	embeddinggemma-300m-q8:latest
Answers	generator.model	qwen3.5:4b
Follow-up rewriting	rewriter.model	qwen3.5:4b
Index-time context (off by default)	enrichment.contextual.model	qwen3.5:4b
Reranker (off by default)	reranker.model	BAAI/bge-reranker-v2-m3

Changing the embedding model changes the vector size: set embedder.dimensions to match, use a new qdrant.collection name or reindex, and adjust the prefixes (query_prefix, document_prefix) that the model expects. The answer, rewriter and enrichment models can be changed without reindexing, except enrichment, which needs a reindex. The demo video and the measured results in the docs used the defaults.

Prerequisites

Tool	Required?	Purpose
Podman + podman-compose	Required for the provided local/Compose flow	Qdrant and optional containerized reranker (rootless; ./scripts/setup_podman_host.sh sets the laptop up)
Go 1.27+	Required	Server + CLI
Python 3.12+	Required for local.sh; sidecars require 3.12+	Local startup config parsing, optional reranker and PDF conversion (the numpy==2.5.2 pin needs 3.12)
Node.js 24 LTS, or 22.x ≥22.12	Required for dashboard	React dashboard and browser tests; Node 25 is unsupported
Ollama	Required	Serves the embedding and answer models (defaults: embeddinggemma-300m-q8, qwen3.5:4b; configurable)

The default text roles use the default answer model with think: false. This keeps their bounded output budgets available for answer text and search rewrites. Each role can set think: true; omitting the setting leaves Ollama's model default in effect.

Quick start

1. Configure your data source

The default has no source directories. Upload documents using the dashboard's paperclip, or edit internal/bootstrap/configuration/config.yaml → documents.paths to enable directory ingestion:

documents:
  mode: "upload-only"     # upload-only | mirror
  paths:
    - "/path/to/documents"

upload-only keeps existing Documents when a source file disappears. Set mode: mirror when the configured directories are the complete corpus; a successful source sweep then removes indexed files missing from those directories. Multipart uploads never trigger mirror deletion.

2. Start everything

./scripts/local.sh

This starts Qdrant and the Go API, starts the host reranker only when enabled, ingests source directories when configured, and blocks on the server. With documents.paths: [], it starts ready for uploads. Reranking is off by default. Run the React dashboard separately:

cd web/dashboard
nvm use # if using nvm; .nvmrc selects Node 24
npm ci
npm run dev

Then open http://localhost:3002. Set DASHBOARD_PORT if that port is also occupied; Vite uses a strict port and will fail clearly instead of silently switching to a different URL.

3. Test search

Upload a document first, then ask about its contents:

curl -X POST localhost:8100/api/v1/turns \
  -H 'content-type: application/json' \
  -d '{"query":"What are the main ideas in the uploaded document?","generate":false}'

The control API uses JSON requests and responses. Live generated answers are delivered by an SSE endpoint in the turn response; the React dashboard handles that stream.

4. Include LLM answer generation

Set generate to true to run answer generation over the retrieved chunks:

curl -X POST localhost:8100/api/v1/turns \
  -H 'content-type: application/json' \
  -d '{"query":"What are the main ideas in the uploaded document?","generate":true}'