Open-source enterprise RAG
ERAG turns your scattered documents into a knowledge engine: connect 24+ sources, ask in plain language, and get cited answers your auditors can trace — from the exact passage to the permission check that allowed it.
Apache-2.0 · runs anywhere, from a laptop to an air-gapped cluster
01Capabilities
Everything between a raw document and a trustworthy answer — built in, not bolted on.
Lexical BM25 written from first principles, dense semantic vectors, reciprocal-rank fusion and a reranking stage — so the right passage wins whether your people search in keywords or in questions.
keyword
semantic
fused + reranked
Access control is evaluated inside every vector and keyword query — never post-filtered, never left to a prompt. What a user can't read can't be retrieved.
Every claim carries a citation. Anti-hallucination modes, a faithfulness self-check, and honest abstention when the corpus doesn't contain the answer.
SharePoint, Confluence, Jira, Slack, GitHub, Google Drive, S3, Azure Blob, GCS, SQL databases and more — on scheduled sync with resumable backfills.
PDFs — including scanned ones via OCR — Office files, images, audio, and YouTube videos: paste a link and the transcript lands in the index with timestamps, so answers cite the exact moment in the video.
Claude, any OpenAI-compatible endpoint, vLLM or Ollama. Or skip the setup: paid plans bundle ERAG-hosted models.
A per-query trace shows retrieval scores, fusion, reranking and the final prompt — exactly why every answer happened.
Golden questions, LLM-judge scoring and concrete tuning recommendations, so retrieval quality is measured — not assumed.
Retention policies, right-to-be-forgotten, PII redaction, audit export to your SIEM, and SCIM user provisioning.
Deduplication, vector quantization, sharded indexes and distributed ingestion keep costs flat as the corpus grows.
02Deployment
One codebase, Apache-2.0 licensed, from first prototype to regulated production.
The zero-dependency dev profile runs with no external services — clone, seed, ask.
Local models via vLLM or Ollama, local storage — nothing leaves the room.
Scale out with sharded indexes and distributed ingestion workers, under your own keys.
03Pricing
One number to think about: how much you index. Unlimited questions, unlimited users.
Estimated monthly cost
$99.50/mo
at $1.99 per GB per month
Includes ERAG-hosted models with 2M AI tokens per GB each month — heavy use continues at $1.50/1M. Or bring your own keys.
Start with 50 GB