Tool Hut
Founder stack playbook
A cited question-answering product over private files and selected web sources, with measurable retrieval quality.
A founder building a knowledge assistant where answers must be grounded in a controlled corpus rather than model memory.
Moderate document volume, text-first sources, permission-aware retrieval, and an evaluation set created before production rollout.
US$0 at prototype volume; typically US$20-50/mo when ingestion or vector hosting leaves free tiers, plus model and embedding usage.
Its ingestion, node metadata, retrievers, and evaluation utilities are centred on data-backed LLM applications.
Skip if: a few SQL queries and a single vector search are clearer than adopting a framework abstraction.
Alternatives: LangChain, Haystack, Custom retrieval pipeline
It turns selected websites into cleaner markdown and handles crawling details that derail an MVP.
Skip if: sources are only uploaded files, robots or licences prohibit crawling, or extraction needs bespoke browser logic.
Alternatives: Apify, Jina AI Reader, Playwright
Metadata filtering, hybrid retrieval support, managed hosting, and a self-host path suit permission-aware RAG.
Skip if: Postgres plus pgvector already meets scale and keeping one database is more valuable than specialist features.
Alternatives: pgvector, Pinecone, Weaviate
S3-compatible object storage keeps originals separate from embeddings and avoids egress charges in common retrieval flows.
Skip if: files already live under reliable access controls in S3, Supabase Storage, Google Drive, or another system of record.
Alternatives: AWS S3, Supabase Storage, Cloudinary
Model choice can be evaluated against grounded-answer quality and cost without changing the retrieval application.
Skip if: documents are sensitive enough that only a contracted provider or local model may receive retrieved passages.
Alternatives: Anthropic API, OpenAI API, Ollama
It supports authenticated upload, streaming answers, citation UI, and server-side provider calls in one deployable app.
Skip if: RAG is an internal API consumed by an existing product rather than a standalone web experience.
Alternatives: React + Vite with Hono, Streamlit, Chainlit
It connects retrieval failures to real user journeys while feature flags support controlled ranking experiments.
Skip if: document sensitivity rules out session capture and a privacy-minimal event system is required.
Alternatives: Langfuse, Plausible, Amplitude
Verification note: Pricing checked 12 July 2026. Cost varies mainly with crawl volume, embedding refreshes, stored vectors, and answer tokens; the estimate excludes OCR-heavy or multimodal corpora.