Embedding + Neural Search

Limitless memory for humans and AI agents.

Enscrive gives humans and AI agents durable, searchable memory. First-class embedding, hybrid and adaptive neural search, and evaluation live behind one /v1 API and a free local CLI. Your agents reach it through the CLI and Enscrive Code today, with the /v1 API opening soon; for people, Enscrive Canvas—a thought processor that remembers—is on the way. Every unit of usage is metered exactly—accounted to the picodollar and verifiable under independent recomputation.

Give your agents memory Install the CLI

Install Enscrive locally · free, no signup

curl -fsSL https://install.enscrive.io/install.sh | sh

Then run enscrive init to bring up the full local stack.

50+ REST Endpoints

8 Embedding Models

1 Region, Pinned by Default

$0 Local CLI

Available now · free, local-first

The full toolkit on your laptop. No account required.

enscrive-cli is a single static binary that gives you the entire Enscrive workflow locally. Same commands, same JSON shapes as the managed API—so the day you decide to ship to a managed plan, nothing changes except the endpoint.

Install in one line, then enscrive init walks you through bootstrapping the local stack—managed mode against api.enscrive.io, or self-managed mode that pulls service binaries from the same signed distribution.

Free for exploration, evaluation, and development — self-managed and non-commercial. Production or commercial use requires a subscription or a commercial license. See License.

Ingest, search, voice tuning, eval runs — all local
Same JSON shapes as /v1 — pipe to jq and ship
Bring your own embedding key (OpenAI, Voyage, Nebius) or self-host BGE
macOS, Linux, x86_64 + arm64 — one installer for all

Install & first search

# install — one line, no account, no signup
curl -fsSL https://install.enscrive.io/install.sh | sh

# bring up the full local stack (managed or self-managed)
enscrive init

# spin up a local corpus and ingest a directory
enscrive corpus create --name knowledge-base --embedding-model openai-text-embedding-3-small
enscrive ingest knowledge-base ./docs

# search it from your shell
enscrive search knowledge-base "how does onboarding work?" --top-k 5

Prefer to call the cloud directly? The same workflows live on the managed /v1 REST API.

The Choice

Months of infrastructure, or minutes to production.

Your AI agents need persistent memory. You can spend months building it from scratch, cobble together commodity APIs, or point your agents at Enscrive and start shipping today.

A purpose-built memory stack engineered for type-safe retrieval from ingestion to result. Your agents need memory. That's our entire focus.

Option A

Build It Yourself

Select embedding providers, deploy vector storage, design chunking, build ranking, write evals.

3–6 months before product code

Option B

Commodity APIs

One API for embeddings, another for vectors, a third for re-ranking. No coherence between them.

Weeks of integration, ongoing glue

Enscrive

Unified by Voices

Embedding, chunking, ranking, evaluation—one platform, one primitive. A Voice is the whole pipeline as one coherent configuration.

Minutes to first search

The unification primitive

Agent Voices: one configuration for the whole pipeline.

A Voice is how Enscrive collapses chunking, embedding, ranking, and evaluation into a single coherent profile. A support bot Voice is different from a research-assistant Voice—but each one is one object: versioned, diffable, eval-gated, promotable across environments. No more keeping four mismatched configs in sync by hand.

Chunking

How documents are split. Baseline (semantic paragraph), Story Beats (LLM-detected narrative boundaries), or Tone Segments (topic shifts).

Embedding

How chunks become vectors. Pick model, dimensions, and resolution. Adaptive Resolution indexes multiple granularities for routed queries.

Retrieval

How results are ranked. Pure vector, hybrid (BM25 + dense), or adaptive. Tunable thresholds, top-k, and dense/sparse alpha per query.

Start from a template, tune from there.

Four built-in templates cover common patterns. Customize or start from scratch.

General Purpose

text-embedding-3-small · 512 tok

Balanced defaults for most use cases. Start here and tune as needed.

High Precision

text-embedding-3-large · 256 tok

Strict matching, low tolerance for noise. High threshold, small chunks.

High Recall

text-embedding-3-small · 1024 tok

Cast a wide net. Lower threshold, larger context, don't miss relevant results.

Conversational

text-embedding-3-small · 384 tok

Optimized for chat-style queries and short-form questions.

Evaluation, first-class · every plan

Prove what you ship. Then ship what you proved.

Most retrieval platforms hand you a vector store and wish you luck. Enscrive treats evaluation as a first-class primitive: every Voice is measurable against a dataset, and every promotion across environments can be gated on the result.

The evals surface is built around a single /v1/evals/{id} control tower—dataset, target, methodology, runs, diagnose—all from one endpoint. Available on every plan, including the free local CLI.

HuggingFace dataset sampling

Stream BeIR, MTEB, or any HuggingFace dataset by URL. Sample stratified subsets for fast iteration; expand to the full corpus when the Voice is dialed in.

Deterministic + LLM-reasoned sampling

Stratified-random with seed for reproducibility, or LLM-reasoned to surface hard queries first. Same eval, two cost profiles.

Standard IR scorers

nDCG@k, Recall@k, Precision@k, MRR — all closed-form Rust, computed the same way for every run. No mystery numbers.

Voice-diff with cost estimator

Before promoting a new Voice, see whether the change is query-only (free) or corpus-invalidating (re-embed, with a $ + wall-clock estimate up front).

Public benchmarks coming. Eval runs against canonical datasets (BEIR/fiqa, scifact, hotpotqa) will publish to /benchmarks—reproducible, dated, pinned to a Voice version.

For Developers

Built for developers. Designed for agents.

A complete /v1 REST API, a free local CLI, and a developer portal with an interactive explorer. Every capability reachable from either client—same shapes, same auth, same Voices.

import requests response = requests.post( "https://us.api.enscrive.io/v1/search", headers={"Authorization": "Bearer ens_sk_..."}, json={ "corpus_id": "knowledge-base", "voice": "high-precision", "query": "How does the onboarding flow work?", "top_k": 5 } ) for result in response.json()["results"]: print(f"[{result['score']:.2f}] {result['text'][:120]}...")

The /v1 API is pre-launch. US-region endpoints (us.api.enscrive.io) are reservable now—install the free CLI above to start today.

Coming Soon

/v1 REST API

50+ endpoints covering the complete lifecycle from corpus creation to eval-gated Voice promotion. Pre-launch—US-region endpoints are reservable now.

Corpora, voices, documents, chunks
Ingest with SSE progress streams
Semantic, hybrid, and adaptive search
Datasets, evals, runs, benchmarks
Jobs, batch sets, staging & commit
Backups, restore, log streaming

Available

enscrive-cli

The full toolkit as a single static binary. Free on your laptop; same commands and shapes as the managed API. SHA256-verified install from a signed distribution.

curl -fsSL https://install.enscrive.io/install.sh | sh
enscrive init bootstraps managed or self-managed mode
Ingest, search, voice tuning, eval runs
Environment switching (dev / staging / prod) built in
Static binary, macOS + Linux, x86_64 + arm64

Available

enscrive-docs

Drop a first-class neural-search /docs endpoint into your existing app. Polished HTML for humans, JSON search for agents.

Markdown directory in, retrieval-native site out
JSON /search + /llms.txt + sitemap
Backed by Enscrive corpora + Voices
Single Rust binary, zero runtime deps
docs.enscrive.io →

Platform

Nine capabilities, one memory stack.

Multi-Provider Embeddings

OpenAI, Voyage.ai, BGE, and Nebius — eight models across four providers. Switch without re-architecting.

Hybrid Search

Dense vectors fused with sparse BM25. Tunable alpha across the dense/sparse spectrum, per query.

LLM-Powered Chunking

Beyond fixed-size splits. Story Beats find narrative boundaries; Tone Segments find topic shifts.

SHA256 Change Detection

Fingerprint-based dedup at ingest. Re-upload unchanged documents at zero cost. Pay only for what changed.

Adaptive Resolution

Multi-granularity embeddings: topic, context, precise. Adaptive mode oversamples then re-ranks.

Multi-Environment

Isolated dev, staging, and production scopes. Promote Voices across environments with eval-gated workflows.

Batch Embedding APIs

Routes through OpenAI, Voyage, and Nebius batch APIs by default. ~50% cost reduction on large ingests.

Staging & Commit

Stage, commit, and revert corpus mutations with an auditable history. No accidental wipes.

Backup & Restore

Incremental backups and point-in-time restore with dry-run preview. Built-in, not a checkbox.

Privacy · Portability · Trust

Your data, your region, your terms.

Region pinning is a contract, not a hint. Every corpus is pinned to a single region for its lifetime — no silent cross-region replication, no "eventual" residency, no surprise moves. Your data is portable on day one, encrypted in transit, and isolated per tenant by design.

Privacy

Single-region pinning, by default

Every tenant — including Enterprise — is pinned to a single region by default (US, Ohio, today). Cross-region deployment is available to Enterprise only by explicit arrangement, with your written sign-off on the cross-regional data-residency terms. No silent replication, no default multi-region footprint.

Privacy

Tenant + environment isolation

Every request is checked against tenant and environment boundaries before it reaches storage. PII (emails, API keys, IPs, bearer tokens) is redacted at the vector pipeline before it ever lands in a log line.

Privacy

Dedicated tenancy (Enterprise)

Enterprise plans run on a dedicated managed instance, single-region by default (US, Ohio, today) — multi-region available by arrangement — with optional VPC peering, BYOK / CMEK, and controls designed and developed to SOC2-aligned standards. Same API, harder boundary.

Portability

Full export & restore

Incremental backups and point-in-time restore are first-class endpoints, not a support ticket. Export corpora, vectors, and Voices as portable artifacts you own and can re-ingest anywhere.

Portability

Open formats & clients

JSONL datasets, BeIR-compatible qrels, plain HTTP + JSON. Both clients (CLI and REST) read and write the same shapes; the local CLI is free for non-commercial dev and evaluation, so leaving the managed plane never strands you.

Security

HMAC + TLS, audited mutations

Constant-time HMAC-SHA256 API key validation, TLS 1.3 in transit, environment-scoped keys, and stage / commit / revert workflows on every corpus change. Auditable by design, not by promise.

Ready to give your agents limitless memory?

Ship on the managed API today, or run the full stack on your laptop with the free CLI— curl -fsSL https://install.enscrive.io/install.sh | sh.

Get Started View Pricing