Document Parsing
PDFs, HTML, and office formats parsed with layout awareness — tables, headings, and reading order preserved.
Hybrid retrieval · Grounded generation · Full observability
Build production-grade AI applications with enterprise RAG
Ingest anything, retrieve with hybrid precision, and stream answers your users can verify — every claim cited, every stage observable.
Trusted by teams at
- Microsoft
- AWS
- Azure
- OpenAI
0.00%uptime SLA
<0msp50 retrieval
0M+documents indexed
veritas — rag shell
Powering retrieval for engineering teams worldwide
- Microsoft
- AWS
- Azure
- OpenAI
- Stripe
- MongoDB
- Redis
- Docker
- Kubernetes
Features
Everything a production RAG stack needs
Every stage of the pipeline is observable, swappable, and built for scale — no glue code required.
OCR
Scanned pages and embedded images pass through OCR automatically, so nothing in your corpus is invisible.
Smart Chunking
Recursive, semantic, and fixed-window strategies with configurable size and overlap — swappable per collection.
Embeddings
Local sentence-transformers or hosted models behind one interface. Change providers with an env var, not a rewrite.
Hybrid Search
Dense vectors and sparse BM25 run in parallel, fused with Reciprocal Rank Fusion for rankings neither achieves alone.
Reranking
Cross-encoder rerankers re-score the candidate set so the context window only carries what actually matters.
Vector Search
Chroma, Pinecone, or Qdrant behind a single store interface — local-first in development, managed at scale.
Streaming
Tokens stream over SSE the moment generation starts, with sources emitted first so citations render instantly.
Citations
Every claim links back to the exact chunk, page, and score it came from. Answers your users can verify.
Guardrails
Retrieved content is treated as data, never instructions — prompt-injection surface minimized by design.
Evaluation
RAGAS-style faithfulness, relevancy, and context precision scored continuously against your live traffic.
Observability
Per-stage latency, token usage, and cache hit rates for every request — Grafana-grade visibility built in.
Pipeline
Every stage, explorable
From upload to citation — click any stage to see its purpose, best practices, library options, and the tradeoffs that matter.
Models
Bring the models you trust
Embeddings, LLMs, and rerankers are pluggable — compare the options and swap providers with a config change, not a rewrite.
Chunking
See exactly how your documents split
Tune chunk size and overlap and watch the segmentation change — the same math the ingestion pipeline runs.
bank-policy.pdf
6 chunks
Savings accounts require a minimum balance of $300, calculated as the average daily balance across each statement cycle. Students and account holders under the age of 25 are exempt from the minimum balance requirement until their 25th birthday. If an account falls below the minimum for two consecutive statement cycles, a monthly maintenance fee of $5 applies until the balance is restored. Checking accounts carry no minimum balance, but overdraft protection requires a linked savings account in good standing. Wire transfers within the same business day are subject to a $25 domestic fee and a $45 international fee, waived for premium tier members. Certificates of deposit may be withdrawn early with a penalty equal to 90 days of accrued interest on terms of one year or less, and 180 days on longer terms. Premium tier membership requires a combined balance of $25,000 across all linked accounts, reviewed quarterly against the average combined balance.
220 chars
40 chars
blended regions belong to two chunks — overlap keeps answers that span a boundary retrievable.
Vector search
Watch retrieval find its neighbors
A 2D window into the embedding space: pick a query and see which chunks the index returns, ranked by similarity.
● query vector · ● nearest chunks · distance in this projection ≈ semantic dissimilarity
Top 5 retrieved
[1] minimum balance rules · chunk 080.94
[2] minimum balance rules · chunk 010.91
[3] minimum balance rules · chunk 000.86
[4] minimum balance rules · chunk 040.85
[5] minimum balance rules · chunk 060.80
Observability
See every request, stage by stage
Latency percentiles, per-stage breakdowns, token throughput, and host telemetry — live for every deployment, no extra wiring.
Evaluation
Quality you can measure, not vibe-check
RAGAS-style scoring runs continuously: faithfulness, relevancy, and context quality tracked across every eval run.
0.00
Faithfulness
Claims supported by retrieved context
0.00
Answer relevancy
Answer addresses the question asked
0.00
Groundedness
No content invented beyond sources
0.00
Context precision
Retrieved chunks that were actually useful
0.00
Context recall
Useful chunks that were actually retrieved
| Run | Dataset | Faithfulness | Relevancy | Precision | Recall | Cost / query | p95 |
|---|---|---|---|---|---|---|---|
| eval-2041 | banking-qa-v3 | 0.96 | 0.93 | 0.89 | 0.91 | $0.0041 | 1.9s |
| eval-2040 | banking-qa-v3 | 0.94 | 0.92 | 0.87 | 0.90 | $0.0043 | 2.1s |
| eval-2039 | support-kb-v1 | 0.91 | 0.90 | 0.84 | 0.88 | $0.0039 | 1.8s |
| eval-2038 | banking-qa-v2 | 0.93 | 0.91 | 0.86 | 0.89 | $0.0046 | 2.3s |
Ship AI your users can actually trust
Start free on your own machine. Scale to millions of documents when you're ready. Every answer cited, every stage observable.
Veritas AI · free tier forever · no credit card required