Migrate to elide toolkit; build engine resource model + orchestrator

Cargo.toml

@@ -3,19 +3,11 @@
 [workspace]
 resolver = "3"
 members = [
+    "./crates/elide-bento",
     "./crates/nvisy-cli",
-    "./crates/nvisy-codec",
-    "./crates/nvisy-context",
     "./crates/nvisy-core",
     "./crates/nvisy-engine",
-    "./crates/nvisy-fake",
-    "./crates/nvisy-llm",
-    "./crates/nvisy-ner",
-    "./crates/nvisy-ocr",
-    "./crates/nvisy-pattern",
     "./crates/nvisy-server",
-    "./crates/nvisy-stt",
-    "./crates/nvisy-toolkit",
 ]
 
 [workspace.package]
@@ -36,19 +28,27 @@ documentation = "https://docs.rs/nvisy-runtime"
 #
 # See for more details: https://github.com/rust-lang/cargo/issues/11329
 
+# Elide toolkit (upstream)
+elide = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-core = { git = "https://github.com/nvisycom/elide", branch = "main" }
+# `elide-ner` + `elide-ocr` are pulled directly only by `elide-bento`, which
+# implements the per-backend traits these crates export. Engine and other
+# consumers reach the same types through `elide::recognition::{ner, ocr}`.
+elide-ner = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-ocr = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-stt = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+
+# Runtime-owned elide extensions
+elide-bento = { path = "./crates/elide-bento", version = "0.1.0" }
+
 # Internal crates
-nvisy-codec = { path = "./crates/nvisy-codec", version = "0.1.0", default-features = false }
-nvisy-context = { path = "./crates/nvisy-context", version = "0.1.0" }
-nvisy-core = { path = "./crates/nvisy-core", version = "0.1.0" }
 nvisy-engine = { path = "./crates/nvisy-engine", version = "0.1.0" }
-nvisy-fake = { path = "./crates/nvisy-fake", version = "0.1.0" }
-nvisy-llm = { path = "./crates/nvisy-llm", version = "0.1.0" }
-nvisy-ner = { path = "./crates/nvisy-ner", version = "0.1.0" }
-nvisy-ocr = { path = "./crates/nvisy-ocr", version = "0.1.0" }
-nvisy-pattern = { path = "./crates/nvisy-pattern", version = "0.1.0" }
 nvisy-server = { path = "./crates/nvisy-server", version = "0.1.0" }
-nvisy-stt = { path = "./crates/nvisy-stt", version = "0.1.0" }
-nvisy-toolkit = { path = "./crates/nvisy-toolkit", version = "0.1.0" }
+
+# Internal crate being deleted: kept here as a path-dep so consumer
+# manifests (nvisy-engine/server/cli) parse during the migration.
+# Leaves the workspace once engine's source no longer imports it.
+nvisy-core = { path = "./crates/nvisy-core", version = "0.1.0" }
 
 # Serialization
 serde = { version = "1.0", features = ["derive"] }

Nvisy.example.toml

@@ -1,13 +1,11 @@
 # Nvisy configuration
 #
-# Copy to Nvisy.toml (or run `make generate-config`), then edit as needed.
-# The server loads Nvisy.toml by default; override with --config or NVISY_CONFIG.
+# Copy to Nvisy.toml (or run `make generate-config`), then edit as
+# needed. The server loads Nvisy.toml by default; override with
+# --config or NVISY_CONFIG.
 #
-# All sections are optional: omitted sections use runtime defaults.
+# All sections are optional: omitted sections use server defaults.
 
-version = "0.1.0"
-
-# Server binding and storage.
 # CLI flags (--host, --port, --data-dir) override these.
 [server]
 host = "0.0.0.0"
@@ -24,107 +22,68 @@ body_limit_mb = 4
 request_timeout = "5m"
 cors = { allowed_origins = ["*"], max_age = "1h" }
 
-# Engine-level resource limits and shared infrastructure.
-[engine]
-
-# Run-level resource limits.
-[engine.limits]
-concurrency = 4              # max parallel documents in flight
-run_timeout = "60s"
-
-# Shared HTTP client for all downstream provider calls.
-[engine.http]
-max_retries = 3
-timeout = "120s"
-connect_timeout = "10s"
-idle_timeout = "90s"
-
-# Extraction registry.
-#
-# Each opted-in `[extraction.*]` section is built once at server
-# startup. Set `enabled = false` to keep the config but skip
-# construction and dispatch. The pipeline `Extraction` step carries
-# per-call flags only.
-
-# Extraction: OCR (optical character recognition, images only).
-[extraction.ocr]
-enabled = true
-
-[extraction.ocr.backend]
-kind = "bento"
-base_url = "http://localhost:3001"
-
-# Extraction: STT (speech-to-text, audio only).
-[extraction.stt]
-enabled = true
-
-[extraction.stt.backend]
-kind = "noop"
-
-# Detection: pattern-based (regex + dictionary + checksum).
-# Optional section — pattern detection runs by default. Include this
-# block to narrow it (named patterns, tag filters) or to disable it
-# entirely with `enabled = false`.
-[detection.pattern]
-enabled = true
-patterns = [] # empty = all built-in patterns
-
-# Detection: NER (named entity recognition).
-[detection.ner]
-enabled = true
-
-[detection.ner.backend]
-kind = "bento"
-base_url = "http://localhost:3000"
-
-# Detection: LLM (large language model).
-[detection.llm]
-enabled = true
-unresolved_policy = "drop" # drop | first-match
-
-[detection.llm.provider]
-kind = "open-ai"
-api_key = "sk-example-replace-me"
-model = "gpt-4o"
-
-[detection.llm.detect]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-# context_window = { ... }   # optional, for chunking large inputs
-# preamble = "..."           # optional, overrides the agent's default system prompt
-
-[detection.llm.verify]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-# context_window = { ... }
-# preamble = "..."
-
-# Detection: VLM (vision-language model, images only).
-[detection.vlm]
-enabled = true
-
-[detection.vlm.provider]
-kind = "open-ai"
-api_key = "sk-example-replace-me"
-model = "gpt-4o"
-
-[detection.vlm.detect]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-
-[detection.vlm.verify]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-
-# Redaction params.
-[redaction]
-confidence_threshold = 0.5
-process_metadata = false
+# The deployment default for AnalyzerParams. Requests carry
+# per-field `analyzer` overrides on top of this default (inherit
+# the slot, replace it, remove it from an optional slot, or patch
+# a list with extend + remove selectors). Omit the whole
+# [analyzer] section and the server falls back to the type-level
+# default — no recognizers, no enrichers, default dedup + scope,
+# empty catalog — useful for tests, not useful in production.
+
+# Caller-asserted scope threaded into every recognizer's context.
+# Add languages and country codes the analyzer should consider
+# (empty = no restriction).
+[analyzer.scope]
+languages = []
+jurisdictions = []
+
+# Deduplication pipeline applied after recognition:
+# calibrate → fuse → resolve → filter.
+[analyzer.deduplication]
+fusion = "max_confidence"        # max_confidence | mean | noisy_or
+resolution = "highest_confidence" # highest_confidence | longest_span
+min_confidence = 0.7              # filter threshold, 0.0..=1.0
+
+# Per-request label catalog. Entries are LabelSchema records:
+# { name, description?, tags?, metadata? }. The server unions
+# every submitted policy's labels with this catalog at request
+# time; entries here are the baseline.
+# [[analyzer.label_catalog]]
+# name = "email_address"
+# description = "RFC 5322 email"
+# tags = ["pii", "contact"]
+
+# Pattern recognizer: at-most-one per analyzer. Loads every
+# shipped pattern + dictionary and wraps the recognizer in
+# elide's context-boost layer.
+[analyzer.recognizers.pattern]
+builtins = true
+context_enhanced = true
+
+# NER recognizer list. Each entry needs a unique `name`; multiple
+# entries run in parallel (e.g. one English model, one Spanish).
+# [[analyzer.recognizers.ner]]
+# name = "default_ner"
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3000", model = "..." }
+
+# LLM recognizer list. Same shape; backend chooses provider.
+# [[analyzer.recognizers.llm]]
+# name = "default_llm"
+# backend = { kind = "mock" }
+
+# Language enricher (writes the document's detected languages
+# into the recognizer context, driving jurisdiction-aware
+# dispatch). At-most-one per analyzer.
+# [analyzer.enrichers.language]
+# min_confidence = 0.6              # 0.0..=1.0; None lets the engine choose
+
+# OCR enricher (image modality only). At-most-one per analyzer.
+# [analyzer.enrichers.ocr]
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3001", model = "..." }
+
+# STT enricher (audio modality only). At-most-one per analyzer.
+# [analyzer.enrichers.stt]
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3002", model = "..." }

crates/elide-bento/Cargo.toml

@@ -0,0 +1,59 @@
+# https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[package]
+name = "elide-bento"
+description = "BentoML-backed NER and OCR backends for the elide toolkit"
+keywords = ["elide", "bento", "bentoml", "ner", "ocr"]
+categories = ["api-bindings"]
+readme = "README.md"
+
+version = { workspace = true }
+rust-version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+publish = { workspace = true }
+
+authors = { workspace = true }
+repository = { workspace = true }
+homepage = { workspace = true }
+documentation = { workspace = true }
+
+[package.metadata.docs.rs]
+all-features = true
+rustdoc-args = ["--cfg", "docsrs"]
+
+[features]
+default = ["ner", "ocr"]
+## Ship the NER backend (implements `elide_ner::backend::NerBackend`
+## against the `nvisy-inference-ner` BentoML service).
+ner = ["dep:elide-ner"]
+## Ship the OCR backend (implements `elide_ocr::OcrBackend`
+## against the `nvisy-inference-ocr` BentoML service).
+ocr = ["dep:elide-ocr", "dep:base64", "elide-core/image"]
+
+[dependencies]
+# Elide toolkit (upstream)
+elide-core = { workspace = true, features = [] }
+elide-ner = { workspace = true, features = [], optional = true }
+elide-ocr = { workspace = true, features = [], optional = true }
+
+# Serialization
+serde = { workspace = true, features = ["derive"] }
+
+# Primitive datatypes
+hipstr = { workspace = true, features = [] }
+
+# Async runtime
+async-trait = { workspace = true, features = [] }
+
+# Error handling
+thiserror = { workspace = true, features = [] }
+
+# Image bytes → base64 (OCR only)
+base64 = { workspace = true, features = [], optional = true }
+
+# BentoML client
+bentoml = { workspace = true, default-features = false, features = ["rustls-tls", "tracing"] }
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["rt", "macros"] }

crates/elide-bento/README.md

@@ -0,0 +1,7 @@
+# elide-bento
+
+Shared BentoML HTTP client wrapper for elide backends.
+
+Per-modality backends (NER, OCR, …) live in their consuming crates
+(`elide-ner`, `elide-ocr`) and pull this crate for the common HTTP
+client, params validation, and error translation.

crates/elide-bento/src/error.rs

@@ -0,0 +1,39 @@
+//! Error translation: `bentoml` errors → [`elide_core::Error`].
+//!
+//! Crate-private — the public API of every backend reports
+//! [`elide_core::Error`]; this enum is the internal seam the
+//! per-route helpers use before bubbling up.
+
+use elide_core::{Error, ErrorKind};
+
+/// Errors surfaced internally by the bento backends.
+///
+/// Two structural categories the consuming crate maps onto
+/// [`ErrorKind`] when bubbling up: transport (HTTP / network /
+/// client construction) and protocol (service answered but the
+/// body did not match the contract — decode error, batch length
+/// mismatch, …).
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum BentoError {
+    /// HTTP / transport failure — client construction, network
+    /// I/O, status-code rejections.
+    #[error("bento transport error: {0}")]
+    Transport(#[from] bentoml::Error),
+    /// Protocol failure — the service answered but the body did not
+    /// match the contract.
+    #[error("bento protocol error: {0}")]
+    Protocol(String),
+}
+
+impl From<BentoError> for Error {
+    /// Map transport to [`ErrorKind::Transport`] and protocol to
+    /// [`ErrorKind::Validation`], carrying the original error as the
+    /// source cause.
+    fn from(err: BentoError) -> Self {
+        let kind = match err {
+            BentoError::Transport(_) => ErrorKind::Transport,
+            BentoError::Protocol(_) => ErrorKind::Validation,
+        };
+        Error::new(kind, err)
+    }
+}

crates/elide-bento/src/lib.rs

@@ -0,0 +1,18 @@
+#![forbid(unsafe_code)]
+#![cfg_attr(docsrs, feature(doc_cfg))]
+#![doc = include_str!("../README.md")]
+
+mod error;
+
+#[cfg(feature = "ner")]
+#[cfg_attr(docsrs, doc(cfg(feature = "ner")))]
+pub mod ner;
+
+#[cfg(feature = "ocr")]
+#[cfg_attr(docsrs, doc(cfg(feature = "ocr")))]
+pub mod ocr;
+
+#[cfg(feature = "ner")]
+pub use self::ner::BentoNer;
+#[cfg(feature = "ocr")]
+pub use self::ocr::BentoOcr;