Cosavu← Products/DataAPI
Your model gets
context,
not fragments.
DataAPI indexes your knowledge base and returns distilled, query-relevant context to your LLM — three retrieval tiers, one endpoint, zero glue code.
Live demo
Ask the knowledge base. Get distilled context, not fragments.
Ask the Cosavu knowledge base
awaiting query input
The problem with RAG
RAG returns chunks.
DataAPI returns context.
Traditional RAG hands your model a bag of raw text fragments — similar-looking but unrelated, often redundant — and hopes it figures out the rest.
DataAPI resolves, re-ranks, deduplicates, and optionally distils the evidence into a single coherent block. The LLM gets what it needs. Nothing more.
Traditional RAG
model receives noise
DataAPI CAR-1.5
model receives signal
Retrieval systems
Three tiers.
One endpoint.
Pass system: "car-0" | "car-1" | "car-1.5" per query. Upgrade a single call without changing your integration.
CAR-0 — Semantic search
Cosavu Store-backed vector search with per-tenant namespace isolation. Fast, reliable, and ideal for most document retrieval workloads where semantic similarity is the primary signal.
- ›Cosavu Store cosine similarity
- ›Per-tenant namespace isolation
- ›HNSW index
- ›Best for: notes, documents, unstructured text
Inside CAR-1
The Engram filter —
seven signals, one score.
Every candidate chunk is scored across semantic similarity, lexical recall, IDF weighting, bigram overlap, exact match, and rank prior. Diversity-aware gating then removes near-duplicate passages before they consume your context budget.
Chunks with zero query-term overlap are penalised 75%. Off-topic passages don't crowd out relevant ones even when the vector space is noisy.
Semantic similarity
35%
Lexical recall
30%
IDF recall
18%
Bigram overlap
12%
Lexical precision
14%
Exact match
6%
Rank prior
2%
Diversity gate
22% novelty min
Capabilities
What DataAPI does
Multi-tenant isolation
Every collection is namespaced to a tenant. Cosavu Index and Cosavu Store are completely isolated — one tenant's data never touches another's.
Ingest anything
Upload PDFs, DOCX, plain text, code files, PPTX, XLSX, and CSV. A universal parser extracts and chunks each file automatically.
Context budget
Returned chunks are trimmed to a configurable token budget using sentence-level extraction — your model never receives a truncated mid-thought.
Safety on every query
A moderation gate runs before retrieval. Harmful queries are blocked at the boundary before any data is accessed.
Get started
Start retrieving with DataAPI.
Get a tenant key and connect your knowledge base in minutes. No infrastructure to manage.