Metadata-Version: 2.4
Name: vincio
Version: 1.8.1
Summary: Vincio: the context engineering platform for AI applications. Compiles prompts, memory, retrieval, tools, schemas, and policies into optimized, validated, observable model-ready context packets.
Project-URL: Homepage, https://github.com/Ohswedd/vincio
Project-URL: Documentation, https://github.com/Ohswedd/vincio/tree/main/docs
Project-URL: Repository, https://github.com/Ohswedd/vincio
Project-URL: Issues, https://github.com/Ohswedd/vincio/issues
Project-URL: Changelog, https://github.com/Ohswedd/vincio/blob/main/CHANGELOG.md
Author: Vincio Contributors
License:                                  Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright 2026 Vincio Contributors
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
License-File: LICENSE
Keywords: agents,ai,context-engineering,evaluation,llm,observability,prompts,rag,structured-output
Classifier: Development Status :: 5 - Production/Stable
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Requires-Python: >=3.11
Requires-Dist: httpx>=0.27
Requires-Dist: pydantic>=2.5
Requires-Dist: pyyaml>=6.0
Requires-Dist: typing-extensions>=4.8
Provides-Extra: all
Requires-Dist: anthropic>=0.34; extra == 'all'
Requires-Dist: boto3>=1.34; extra == 'all'
Requires-Dist: chromadb>=0.5; extra == 'all'
Requires-Dist: duckdb>=0.10; extra == 'all'
Requires-Dist: elasticsearch>=8.0; extra == 'all'
Requires-Dist: fastapi>=0.110; extra == 'all'
Requires-Dist: google-cloud-storage>=2.16; extra == 'all'
Requires-Dist: google-genai>=0.3; extra == 'all'
Requires-Dist: lancedb>=0.10; extra == 'all'
Requires-Dist: langchain-core>=0.3; extra == 'all'
Requires-Dist: llama-index-core>=0.11; extra == 'all'
Requires-Dist: mistralai>=1.0; extra == 'all'
Requires-Dist: neo4j>=5.0; extra == 'all'
Requires-Dist: networkx>=3.0; extra == 'all'
Requires-Dist: openai>=1.40; extra == 'all'
Requires-Dist: opensearch-py>=2.4; extra == 'all'
Requires-Dist: opentelemetry-api>=1.24; extra == 'all'
Requires-Dist: opentelemetry-sdk>=1.24; extra == 'all'
Requires-Dist: pdfplumber>=0.11; extra == 'all'
Requires-Dist: pgvector>=0.2; extra == 'all'
Requires-Dist: pinecone>=5.0; extra == 'all'
Requires-Dist: psycopg[binary]>=3.1; extra == 'all'
Requires-Dist: pydantic-settings>=2.2; extra == 'all'
Requires-Dist: pymilvus>=2.4; extra == 'all'
Requires-Dist: pypdf>=4.0; extra == 'all'
Requires-Dist: python-docx>=1.1; extra == 'all'
Requires-Dist: pyvespa>=0.40; extra == 'all'
Requires-Dist: qdrant-client>=1.9; extra == 'all'
Requires-Dist: rank-bm25>=0.2; extra == 'all'
Requires-Dist: redis>=5.0; extra == 'all'
Requires-Dist: tiktoken>=0.7; extra == 'all'
Requires-Dist: uvicorn>=0.29; extra == 'all'
Requires-Dist: weaviate-client>=4.5; extra == 'all'
Requires-Dist: websockets>=12.0; extra == 'all'
Provides-Extra: anthropic
Requires-Dist: anthropic>=0.34; extra == 'anthropic'
Provides-Extra: chroma
Requires-Dist: chromadb>=0.5; extra == 'chroma'
Provides-Extra: dev
Requires-Dist: mypy>=1.9; extra == 'dev'
Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
Requires-Dist: pytest>=8.0; extra == 'dev'
Requires-Dist: ruff>=0.4; extra == 'dev'
Provides-Extra: docx
Requires-Dist: python-docx>=1.1; extra == 'docx'
Provides-Extra: duckdb
Requires-Dist: duckdb>=0.10; extra == 'duckdb'
Provides-Extra: elasticsearch
Requires-Dist: elasticsearch>=8.0; extra == 'elasticsearch'
Provides-Extra: gcs
Requires-Dist: google-cloud-storage>=2.16; extra == 'gcs'
Provides-Extra: google
Requires-Dist: google-genai>=0.3; extra == 'google'
Provides-Extra: graph
Requires-Dist: neo4j>=5.0; extra == 'graph'
Requires-Dist: networkx>=3.0; extra == 'graph'
Provides-Extra: lancedb
Requires-Dist: lancedb>=0.10; extra == 'lancedb'
Provides-Extra: langchain
Requires-Dist: langchain-core>=0.3; extra == 'langchain'
Provides-Extra: llamaindex
Requires-Dist: llama-index-core>=0.11; extra == 'llamaindex'
Provides-Extra: milvus
Requires-Dist: pymilvus>=2.4; extra == 'milvus'
Provides-Extra: mistral
Requires-Dist: mistralai>=1.0; extra == 'mistral'
Provides-Extra: openai
Requires-Dist: openai>=1.40; extra == 'openai'
Provides-Extra: opensearch
Requires-Dist: opensearch-py>=2.4; extra == 'opensearch'
Provides-Extra: otel
Requires-Dist: opentelemetry-api>=1.24; extra == 'otel'
Requires-Dist: opentelemetry-sdk>=1.24; extra == 'otel'
Provides-Extra: pdf
Requires-Dist: pypdf>=4.0; extra == 'pdf'
Provides-Extra: pdf-layout
Requires-Dist: pdfplumber>=0.11; extra == 'pdf-layout'
Provides-Extra: pinecone
Requires-Dist: pinecone>=5.0; extra == 'pinecone'
Provides-Extra: postgres
Requires-Dist: pgvector>=0.2; extra == 'postgres'
Requires-Dist: psycopg[binary]>=3.1; extra == 'postgres'
Provides-Extra: realtime
Requires-Dist: websockets>=12.0; extra == 'realtime'
Provides-Extra: redis
Requires-Dist: redis>=5.0; extra == 'redis'
Provides-Extra: retrieval
Requires-Dist: qdrant-client>=1.9; extra == 'retrieval'
Requires-Dist: rank-bm25>=0.2; extra == 'retrieval'
Provides-Extra: s3
Requires-Dist: boto3>=1.34; extra == 's3'
Provides-Extra: server
Requires-Dist: fastapi>=0.110; extra == 'server'
Requires-Dist: pydantic-settings>=2.2; extra == 'server'
Requires-Dist: uvicorn>=0.29; extra == 'server'
Provides-Extra: tokenizers
Requires-Dist: tiktoken>=0.7; extra == 'tokenizers'
Provides-Extra: vespa
Requires-Dist: pyvespa>=0.40; extra == 'vespa'
Provides-Extra: weaviate
Requires-Dist: weaviate-client>=4.5; extra == 'weaviate'
Description-Content-Type: text/markdown

<p align="center">
  <img src="assets/banner.svg" alt="Vincio — the context engineering platform for AI applications" width="660">
</p>

<p align="center">
  <em>The scarce resource is not the model. It is the context you feed it.</em>
</p>

<p align="center">
  <a href="https://pypi.org/project/vincio/"><img src="https://img.shields.io/pypi/v/vincio?color=B98B2E" alt="PyPI version"></a>
  <a href="https://github.com/Ohswedd/vincio/actions/workflows/ci.yml"><img src="https://github.com/Ohswedd/vincio/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
  <img src="https://img.shields.io/pypi/pyversions/vincio?logo=python&logoColor=white" alt="Python 3.11+">
  <img src="https://img.shields.io/badge/license-Apache%202.0-4C6EF5" alt="Apache 2.0">
  <img src="https://img.shields.io/badge/tests-919%20passing-2ea44f" alt="919 tests passing">
  <img src="https://img.shields.io/badge/lint-ruff-D7FF64" alt="Ruff">
  <img src="https://img.shields.io/badge/typed-pydantic%20v2-E92063" alt="Pydantic v2">
  <img src="https://img.shields.io/badge/offline-first-555" alt="Offline-first">
</p>

---

**Vincio** is a Python platform for building **context-engineered** AI applications. It compiles
prompts, memory, retrieval, tools, schemas, and policies into optimized, testable, observable,
provider-neutral **context packets** — then validates and evaluates every output.

Most LLM frameworks help you call a model. Vincio governs the *boundary* between your application
state and the model: what evidence is selected, how it is scored and budgeted, how it is rendered
for cache reuse, and how the result is validated, measured, and traced. Named for **Leonardo da
Vinci** — engineering and craft in equal measure.

```text
Raw Input → Normalization → Objective Detection → Memory Selection
→ Retrieval Planning → Evidence Retrieval → Ranking + Distillation
→ Tool Planning → Context Compilation → Model Execution
→ Parsing + Validation → Evaluation + Guardrails → Trace + Learning Loop
```

## Contents

[Why Vincio](#why-vincio) · [Install](#install) · [60-second quickstart](#60-second-quickstart) ·
[Features](#features) · [Benchmarks](#benchmarks) · [Comparison](#how-vincio-compares) ·
[Use cases](#use-cases) · [Examples](#more-examples) · [CLI](#command-line) ·
[Architecture](#architecture) · [Roadmap](#roadmap) · [Documentation](#documentation)

## Why Vincio

Teams ship a prompt, watch it work, then spend months fighting everything around it: context that
overflows the window, retrieved chunks that contradict each other, outputs that fail to parse,
silent quality regressions, untraceable costs, and prompt-injection risk. These are not model
problems — they are **context** problems.

Vincio treats context as a compiled artifact with a clear contract:

- **Deterministic where it matters.** Security, permissions, and validation are enforced in code —
  never gated on model output. The same input compiles to the same packet.
- **Measured, not asserted.** Every run is traced and costed; every change can be gated by an eval
  suite before it ships.
- **Provider-neutral.** OpenAI, Anthropic, Google, Mistral, any OpenAI-compatible endpoint, or a
  deterministic offline mock — behind one interface.
- **One coherent model** from input to output, instead of a bag of loosely-coupled utilities.

## Install

```bash
pip install vincio                  # core — runs fully offline with the mock provider
pip install "vincio[openai]"        # + OpenAI provider
pip install "vincio[anthropic]"     # + Anthropic provider
pip install "vincio[chroma]"        # + a vector store (also: pinecone, lancedb, postgres,
                                    #   weaviate, milvus, elasticsearch, opensearch, vespa)
pip install "vincio[realtime]"      # + voice/realtime sessions (OpenAI Realtime, Gemini Live)
pip install "vincio[langchain]"     # + LangChain interop export (also: llamaindex)
pip install "vincio[all]"           # every optional integration
```

Python 3.11+. Core dependencies are just `pydantic`, `httpx`, `pyyaml`, and `typing-extensions`;
every heavy integration (vector stores, OCR, server, OpenTelemetry, …) is an opt-in extra.

## 60-second quickstart

```python
from vincio import ContextApp

app = ContextApp(name="docs_qa")
app.add_source("docs", path="./docs", retrieval="hybrid")
app.set_policy("answer_only_from_sources", True)

result = app.run("How do I configure SSO?")
print(result.output)      # the grounded answer
print(result.citations)   # evidence the answer actually cited
print(result.trace_id)    # every run produces a full trace
print(result.cost_usd)    # …and a cost
```

No API key? It runs offline out of the box on a deterministic mock provider that emits
schema-valid output — so your whole pipeline (retrieval, validation, evals, traces) runs for real
in CI.

### Typed output

```python
from pydantic import BaseModel
from vincio import ContextApp

class TicketClassification(BaseModel):
    label: str
    confidence: float
    reason: str

app = ContextApp(name="triage", output_schema=TicketClassification)
result = app.run("The dashboard crashes after login")

result.output.label        # → a validated TicketClassification instance
```

### Agents with tools and memory

```python
app = ContextApp(name="support_refunds", output_schema=RefundDecision)
app.add_memory(scope="user", strategy="semantic")
app.add_tool("billing_lookup", permissions=["billing:read"])
app.add_tool("refund_create", permissions=["billing:write"], approval_required=True)

agent = app.agent(max_steps=6)
result = agent.run("Customer asks for a refund on invoice INV-123.")
```

### Multi-agent crews and durable graphs

```python
from vincio.agents import interrupt

crew = app.crew(members=[
    {"name": "researcher", "goal": "gather the numbers", "keywords": ["find"]},
    {"name": "writer", "goal": "draft the recommendation"},
])
result = crew.run("Explain the Q3 refund trend")   # bounded, traced, blackboard-shared

graph = app.graph("review")                        # checkpointed in your own store
graph.add_node("analyze", analyze)
graph.add_node("approve", lambda s: {"ok": interrupt(s, "proceed?")})
graph.add_edge("analyze", "approve")
flow = graph.compile()
paused = flow.invoke({"doc": "msa.pdf"})           # pauses at the human gate
done = flow.resume(paused.thread_id, value=True)   # later — even after a restart
```

### Reliability as a guarantee

```python
from vincio import Signature, InputField, OutputField

class Triage(Signature):
    """Classify a support ticket."""
    ticket: str = InputField(desc="the raw ticket text")
    label: str = OutputField(desc="bug | billing | feature | other")
    confidence: float = OutputField()

result = app.predictor(Triage)(ticket="The export button 500s")  # typed + validated

app.add_rail(name="no_leaks", kind="safety", detectors=["pii", "secrets"], action="redact")
app.add_rail(name="on_topic", kind="topic", direction="input", blocked_topics=["legal advice"])
app.enable_self_correction(max_cycles=2, max_cost_usd=0.05)      # facts never invented
app.add_output_schema(BugReport, keywords=["bug", "crash"])       # multi-schema routing

async for event in app.astream("Extract the invoice"):
    if event.type == "partial_output" and event.valid_prefix is False:
        break   # streaming validation: stop paying for a doomed answer
```

### Evaluation as a gate

```python
from vincio.evals import Dataset, EvalRunner

dataset = Dataset.load("golden/support_triage.jsonl")
report = EvalRunner(app).run(dataset)
report.print_summary()     # groundedness, citation accuracy, schema validity, cost — with CI exit codes
```

### Interoperate: MCP, A2A, Skills

```python
# Consume an MCP server — its tools run through the permissioned, sandboxed,
# audited runtime; its resources become cited evidence.
app.add_mcp_server("weather", command=["python", "weather_server.py"])

# Load portable SKILL.md procedural knowledge (progressive disclosure).
app.add_skill("skills/pdf-invoice")

# Expose your app over the protocols — one ContextApp, both consumer and provider.
mcp_server = app.serve_mcp()                       # serve tools/resources/prompts
a2a_server = app.serve_a2a(crew, name="research")  # Agent Card + task lifecycle

# One portable reasoning knob across providers (thinking tokens are billed).
from vincio.core.types import RunConfig
app.run("Plan the migration", config=RunConfig(reasoning_effort="high"))
```

## Features

Vincio is organized into composable subsystems. Use the high-level `ContextApp` runtime, or reach
for any engine directly.

| Subsystem | What it does |
|---|---|
| **Prompt compiler** | Typed prompt ASTs with `${variables}`, lint rules, cache-aware stable-prefix layout, versioning, hashing, diffing, variant generation. |
| **Context compiler** | Scores every candidate (relevance, novelty, authority, freshness, provenance, token cost, leakage risk), deduplicates, resolves conflicts, compresses, and packs to a token budget — with an *excluded-context report* explaining every omission. |
| **Retrieval (RAG)** | BM25 + dense + learned-sparse (SPLADE-style) + late-interaction (ColBERT-style MaxSim with PLAID-style compression) fused in one weighted RRF; query understanding (HyDE, multi-query, decomposition, step-back); sentence-window, parent-document/auto-merging, and contextual chunking; GraphRAG with community summaries and global/local routing; live indexes (upsert/TTL/migrations); entity-graph, multi-hop, and reasoning retrieval; **Matryoshka (MRL) dimension truncation**, **contextual (Voyage context-3)** and **unified text+image multimodal (Cohere v4 / Voyage)** embedders, and query-vs-document input-type hints behind one `build_embedder`; citations. |
| **Memory** | Layered (session → episodic → semantic → tenant → graph) with a guarded write pipeline, confidence decay, contradiction resolution, and privacy scoping; `remember`/`recall` personalization over user/agent/session scopes, hybrid vector+graph recall, episodic→semantic consolidation with provenance, TTL + importance-weighted retention, audited GDPR-style edit/forget/export/erase, and a CI-gated memory eval harness. |
| **Tools** | Permissioned registry (RBAC scopes + ABAC rules), schema derivation from type hints, a resource-limited sandbox (timeout, output caps, scrubbed env, POSIX CPU/memory/fd `setrlimit`), reliability scoring, idempotent write-action guardrails with approval callbacks. |
| **Agents** | Bounded DAG execution with planners (direct / static / dynamic / ReAct / plan-and-execute), critics, validators, human gates, and hard budget enforcement. |
| **Orchestration** | Multi-agent crews — roles, delegation, and a shared versioned blackboard — with per-agent budget shares and guaranteed termination; durable stateful graphs with checkpoints on your storage, resume, edit-and-resume, and time-travel forks; first-class human-in-the-loop interrupts; a declarative `compose`/pipe API with streaming node events; runtime backends exporting to LangGraph and the OpenAI Agents SDK. |
| **Workflows** | Deterministic DAGs with retries, branching, parallelism, compensation, and approval gates that pause the run and resume without re-executing finished steps. |
| **Structured output** | Pydantic output contracts, provider-native constrained decoding with strict schema sanitization (robust-parser fallback everywhere else), streaming validation with mid-stream early abort, DSPy-style typed signatures (`Signature` / `Predict`) that feed the optimizer, bounded self-correcting loops with cost ceilings, multi-schema routing by task or content, and **principled repair that fixes structure only — never invents facts**. |
| **Evaluation** | Golden JSONL datasets, 30+ task / grounding / quality / safety / conversational / **trajectory & tool-use** / retrieval / operational metrics (faithfulness, answer relevance, hallucination with strict number checks, toxicity, bias, summarization, knowledge retention, tool-call accuracy/F1, goal accuracy, plan adherence, step efficiency), deterministic / model / G-Eval judges with calibration, synthetic dataset generation with provenance, red-teaming judged by the security detectors, experiment tracking with statistical significance, regression gates, and baseline-diff reports — plus a `pytest` plugin (`assert_eval` / `assert_grounded`, packet/trace snapshots). |
| **Agentic eval & continuous quality** | Score *how* a run reached its answer, not just the text: trajectory & tool-use metrics over a `Trajectory` projected from any crew / graph / trace (no re-instrumentation); a deterministic multi-turn `Simulator`; **online evaluation** on a sampled slice of live traffic (score time series, off the hot path); **drift detection** (score + embedding-distribution) raising a `drift.detected` event; human annotation with **Cohen's-κ** judge calibration; production A/B with cost + significance per variant. Every metric doubles as a runtime guardrail (`add_metric_rail`) and optimizer fitness term. |
| **Optimization** | Prompt / context / routing / cache search driven by an eval-fitness function, with safety-gated promotion that blocks any candidate regressing schema validity or safety. |
| **The closed loop** | One continuous, reproducible cycle — trace → dataset → eval → optimize → promote (`ImprovementLoop` / `vincio loop run`): production traces become datasets, the gated optimizer searches, and the winner lands in the prompt registry tagged, eval-linked, applied live, and audited. Plus: grounded auto-memory from runs, eval-driven retrieval feedback (gated fusion/reranker tuning, chunking recommendations), cost/quality Pareto frontiers with knee-point selection, learned per-task budget allocation, and hill-climb/annealing search strategies — every signal flowing through one packet, ledger, and trace. |
| **Reflective optimization & the flywheel** | A GEPA-style `ReflectiveOptimizer` that reads eval failures, reflects on why a prompt lost, and proposes targeted edits, evolving a Pareto frontier under a hard rollout budget (plus MIPRO joint instruction+example proposal); a **distillation flywheel** (`app.export_training_set` / `vincio distill`) that curates grounded production traces into provider-ready fine-tuning JSONL and gates a cheaper student into the routing cascade only when it holds quality; **learned prompt compression** (`LLMLinguaCompressor`) as a faithfulness-gated compiler pass; and reflective calibration of the optimizer's own LLM judge against κ-validated labels. |
| **Observability** | Every run yields a full trace span tree with sessions, threaded runs, user feedback, and eval scores on spans; JSONL and OpenTelemetry exporters (GenAI semantic conventions); a local viewer (TUI + self-contained static HTML export + visual trace diff); traces become eval datasets in one command; a versioned prompt registry with tags, diffs, rollback, and eval links; per-run cost tracking. |
| **Security** | Deterministic PII / secret detection and redaction (with non-English locale packs for France/Germany/Spain/India/Singapore/Brazil/UK), prompt-injection defense, **authority/provenance RAG-poisoning detection** on retrieved evidence, programmable input/output rails (topic / format / safety / custom) in the deterministic policy engine, RBAC / ABAC, tenant isolation, and a hash-chained audit log with offline tamper verification (`vincio audit verify`) — all documented in a [threat model](docs/security/threat-model.md) and shipped with SBOM + SLSA provenance attestations. |
| **Governance & compliance** | Evidence generated from the running system, as files you own: machine-readable **model & system cards** (`app.model_card` / `system_card`), a **compliance coverage matrix** across OWASP LLM Top 10 (2025) / OWASP Agentic / NIST AI RMF / MITRE ATLAS backed by red-team and eval evidence (`app.compliance_report`), an **AI-BOM** with SHA-256 model-hash verification (`app.aibom`), EU AI Act **synthetic-content marking** (optionally signed + verifiable), **data lineage** with right-to-erasure-by-source (`app.erase_source`), **data-residency-aware** egress refusal that infers the region from a pinned endpoint (`app.set_residency`), and the non-English **token tax** surfaced per language/tenant — see the [governance guide](docs/guides/governance.md). |
| **Storage** | Pluggable metadata (in-memory / SQLite / Postgres), blob, analytics (DuckDB), vector (Qdrant / pgvector / Chroma / Pinecone / LanceDB / Weaviate / Milvus / Elasticsearch / OpenSearch / Vespa behind one `build_vector_index` factory), and graph (Neo4j) backends. |
| **Providers** | OpenAI (Chat Completions + Responses API), Anthropic, Google, Mistral, any OpenAI-compatible endpoint (with hosted-gateway presets: groq, together, fireworks, openrouter, deepseek, perplexity, xai, nvidia), and a deterministic offline mock — all async-first with sync wrappers, pooled transport, retries, and **capability-aware failover** (refuses a substitution that can't serve the request), with in-flight request coalescing. A data-driven `ModelRegistry` (capabilities, pricing, lifecycle) is the single source the cost table, capability guards, and rotation all read from. Unified reasoning control (`reasoning_effort` / thinking budget) maps across OpenAI/Anthropic/Gemini, with thinking tokens recorded and billed. Opt-in **voice/realtime** sessions (OpenAI Realtime, Gemini Live) via `vincio.realtime` — VAD, interruption, and in-session tool calls through the permissioned runtime. |
| **Protocols & interoperability** | Speaks the standards in-process: **MCP** client *and* server (stdio / Streamable HTTP / in-process) — MCP tools run through the permissioned, sandboxed, audited, budgeted runtime; resources become cited evidence. **A2A** agent-to-agent — expose a crew/graph as an Agent Card + task lifecycle, and reach remote agents as bounded, traced crew delegates. **Agent Skills** — `SKILL.md` with progressive disclosure, bundled scripts as sandboxed tools. All via `app.add_mcp_server` / `serve_mcp` / `serve_a2a` / `add_skill` (experimental, since 1.1). |
| **Performance** | End-to-end streaming (`astream` + SSE) with incremental partial-JSON output, concurrent retrieval/memory/tool fan-out with cancellation propagation and hard latency deadlines, content-addressed compile/chunk/embedding caches, zero-copy (slim) context packets, and CI-gated VincioBench performance budgets. |
| **Cost & reliability (FinOps)** | Production-traffic resilience in-process, not a proxy hop: **batch execution** at ~50% cost (OpenAI Batch + Anthropic Message Batches + Google/Vertex batch, `app.batch`), **circuit breaking** + **health-aware failover** and **key pooling** with RPM+TPM rate limiting (retry → fallback → circuit-break), **runtime model cascades** (start cheap, escalate on low confidence, `app.use_cascade`), **cost attribution** by tenant/feature (`app.cost_report` / `vincio cost report`) with enforced **budget SLOs** (cap / degrade / queue-to-batch + `cost.anomaly`), provider-aware **prompt caching** with TTL choice and cache-hit telemetry, and **incremental** (content-hash) + **sharded** indexing at scale. |
| **Provider/model rotation & swap regression** | The migration safety net for the riskiest production change. A registry-backed **`Router`** (`app.use_router`) picks the cheapest / fastest / least-busy *capable* model per request and downgrades to a per-request budget. A **`SwapGate`** (`app.gate_swap` / `vincio providers regress`) replays golden traces and runs an eval + cost + latency + behavioral diff with statistical significance, PASS/FAILing the swap; **model-swap regression** (`app.swap_regression` / `vincio eval regress`) swaps *only* the model and reports per-metric significance, the cost/latency trade, and the worst-regressed slices, with `repeats`/flake quarantine. A **shadow provider** and a capped **canary** qualify a candidate on live traffic with automatic rollback (`app.shadow` / `app.canary`), and a **lifecycle watcher** (`app.watch_lifecycle` / `vincio providers lifecycle`) proposes migrations off deprecated models (experimental, since 1.8). |
| **Connectors** | Pluggable data connectors — web, GitHub, SQL, S3, GCS, Notion, Confluence, Slack, plus custom via `register_connector` — feeding the document engine with full provenance: `app.add_source("kb", connector=connect("github", repo="acme/handbook"))`. |
| **Integrations & DX** | LangChain + LlamaIndex interop (`vincio.interop`) for tools, retrievers, loaders, and embeddings — both directions, duck-typed `from_*` (no heavy import); hosted rerankers/embedders (Cohere/Jina/Voyage, httpx-only) behind `build_reranker`/`build_embedder`; opt-in domain packs (support, engineering, finance, legal) via `app.use_pack(...)`; `vincio init` templates (rag/agent/eval) with a typed `vincio.yaml` JSON Schema for editor completion; notebook reprs (`enable_rich_reprs`) and an interactive `vincio tui` inspector. |
| **Stability & guarantees** | [Semantic Versioning](https://semver.org/spec/v2.0.0.html) on a frozen public surface (`vincio.__all__`) with a mechanical [deprecation policy](docs/reference/stability.md) (`@deprecated` / `@experimental` / `stability_of`); published performance & quality [SLOs](docs/reference/slo.md) held by at-least-as-strict VincioBench budgets; CycloneDX SBOM + SLSA build-provenance attestations on every release. |

Every extension point — providers, metrics, chunkers, rerankers, judges, validators, tools — accepts
your own implementation via a registry.

## Benchmarks

**VincioBench** ships in `benchmarks/` and runs fully offline (deterministic provider + deterministic
metrics) so results are reproducible. Each family compares the Vincio pipeline against a naive
baseline. Representative results on the bundled reference corpus:

| Family | Metric | Vincio | Naive baseline |
|---|---|--:|--:|
| **Context compression** | evidence tokens for the same task | **216** | 1,175 (stuff-everything) |
| | → token reduction | **−81.6%** | — |
| **Output recovery** | malformed model outputs successfully parsed | **5 / 5** | 3 / 5 (`json.loads`) |
| **Security** | prompt-injection detection rate | **100%** | — |
| | injection false-positive rate | **0%** | — |
| | PII coverage | **100%** | — |
| **Retrieval** | recall@3 / MRR (known-answer corpus) | **1.00 / 1.00** | — |
| | per-mode recall@3 (sparse · late-interaction · PLAID · hybrid_full) | **1.00 each** | — |
| **Memory** | preference recall · contradiction supersede · tenant isolation | **pass** | — |
| **Tools** | runtime overhead, p50 | **0.02 ms** | — |
| **Agents** | adversarial infinite-loop model | **bounded** (budget) | unbounded |
| **Orchestration** | crew over-budget termination · delegation recorded | **pass** | — |
| | graph interrupt→resume and fork-replay vs straight run | **identical state** | — |
| **Evals** | metric agreement on labeled examples | **100%** | — |
| | red-team detector coverage · guarded attack success | **100% · 0%** | naive target: 85% attacks succeed |
| | A/B significance (real shift detected / null ignored) | **pass** | — |
| **Reliability** | invalid output detected mid-stream → tokens saved | **98%** | 0% (validate at end) |
| | self-correction recovery rate (bounded cycles) | **3 / 3** | — |
| | rail catch rate · false positives on clean text | **100% · 0** | — |
| | schema routing / classification accuracy | **100%** | — |
| **Closed loop** | loop promotion fires · deterministic · gates block regressions | **pass** | — |
| | grounded facts written · ungrounded excluded | **pass** | — |
| | retrieval feedback (noisy index corrected · healthy index untouched) | **pass** | — |
| | Pareto front excludes dominated · knee balanced · learned budgets promote | **pass** | — |
| **Protocols** | MCP tool schema fidelity · resource provenance · round-trip | **1.00 · pass · pass** | thin adapter |
| | A2A budget-bounded delegation terminates | **pass** | — |
| | Agent-Skill progressive-disclosure token savings (off-topic) | **100%** | — |
| **Cost & reliability (scale)** | batch reconciliation by custom id · partial failures surfaced | **pass · pass** | dropped silently |
| | circuit opens on systemic failure → half-open recovers · failover steers healthy | **pass** | one slow timeout/request |
| | prompt-cache hit rate (warm stable prefix) · cost-attribution accuracy | **72% · 100%** | — |
| | cascade savings vs always-strong (cheap-first, escalate on low confidence) | **−70%** | — |
| | canary auto-rolls-back under load → serves primary after | **pass** | serves the regression |
| **Provider/model rotation & swap** | router picks cheapest capable · budget downgrade · capability filter | **pass** | routes to incapable/pricier |
| | capability guard skips incapable failover · lifecycle error classified (rotate now) | **pass** | silent wrong answer / buried 404 |
| | swap gate blocks a regression · passes a safe swap (significance) · replay-diff fidelity | **pass** | swap on one noisy run |
| | Google/Vertex batch parity (half-cost) | **pass** | — |
| **Reflective optimization & flywheel** | reflective search beats baseline within rollout budget · deterministic | **pass** | blind mutation |
| | distillation exports grounded-only · gates student on quality hold | **pass** | trains on hallucinations |
| | learned compression preserves cited facts under faithfulness gate | **pass** | drops evidence |
| **Governance** | card/AI-BOM completeness · framework-mapping coverage | **pass · 79%** | — |
| | erasure correctness (chunks removed = lineage) · audited | **pass** | — |
| | multilingual PII recall · RAG-poisoning detection (FP rate) | **100% · 100% (0%)** | English-only |

> **Honest by design.** These numbers come from a small, synthetic offline corpus and are meant to
> demonstrate the mechanisms, not to be quoted as universal gains. The context-compression
> hypothesis (a 20–40% reduction target) is *measured* per run, and VincioBench reports whether it
> was met on your data. Run `python benchmarks/vinciobench.py` against your own corpus — and trust
> only what that prints. See [`benchmarks/README.md`](benchmarks/README.md).

## How Vincio compares

Each ecosystem below is broad and capable in its own focus area. The table reflects **built-in,
in-library** capabilities — not what is reachable by bolting on a separate product or SaaS.

| Capability | **Vincio** | LangChain | LlamaIndex | DSPy | Ragas |
|---|:--:|:--:|:--:|:--:|:--:|
| Scored, budgeted **context compiler** | ✅ | ➖ | ➖ | ❌ | ❌ |
| Typed prompt **AST + lint + cache layout** | ✅ | ❌ | ❌ | ➖ | ❌ |
| Hybrid (BM25 + dense) **RAG** | ✅ | ✅ | ✅ | ❌ | ❌ |
| **Sparse + late-interaction + GraphRAG** in one fusion | ✅ | ➖ | ➖ | ❌ | ❌ |
| Layered **memory** (decay, conflicts, scopes) | ✅ | ➖ | ➖ | ❌ | ❌ |
| **Permissioned** tool registry (RBAC/ABAC) | ✅ | ❌ | ❌ | ❌ | ❌ |
| Bounded **agents** + deterministic workflows | ✅ | ✅ | ➖ | ➖ | ❌ |
| **Durable graphs** (checkpoint / resume / time-travel) + bounded crews | ✅ | ➖ | ❌ | ❌ | ❌ |
| Structured output + **structure-only repair** | ✅ | ➖ | ➖ | ✅ | ❌ |
| Built-in **evals + CI gates** | ✅ | ➖ | ➖ | ➖ | ✅ |
| **pytest assertions + red-teaming + synthetic data** | ✅ | ❌ | ❌ | ❌ | ➖ |
| Eval-driven **optimization** (gated promotion) | ✅ | ❌ | ❌ | ✅ | ❌ |
| Native **tracing + cost**, no account needed | ✅ | ➖ | ➖ | ❌ | ❌ |
| **Sessions, feedback, prompt registry, trace viewer** in-process | ✅ | ➖ | ❌ | ❌ | ❌ |
| **Deterministic security** (PII / injection / audit) | ✅ | ❌ | ❌ | ❌ | ❌ |
| **MCP** client *and* server + **A2A** + **Agent Skills** | ✅ | ➖ | ➖ | ➖ | ❌ |
| **In-process FinOps**: batch · circuit-break · cascades · cost attribution + budgets | ✅ | ❌ | ❌ | ❌ | ❌ |
| **Capability-aware rotation + gated swap regression** (router · swap gate · canary) | ✅ | ➖ | ❌ | ❌ | ❌ |
| **Governance evidence**: cards · OWASP/NIST/MITRE mapping · AI-BOM · erasure · residency | ✅ | ❌ | ❌ | ❌ | ❌ |

<sub>✅ first-class in-library · ➖ partial or via a separate add-on/SaaS · ❌ not a focus. Reflects
mid-2026; ecosystems evolve. Vincio is built to *interoperate* — it speaks MCP (client *and* server),
A2A, and Agent Skills in-process, `vincio.interop` brings LangChain and LlamaIndex tools, retrievers,
loaders, and embeddings in (and hands Vincio's back), and you can point at any OpenAI-compatible model
and the vector store you already run. See the
[migration guides](docs/guides/migrate-from-langchain.md), the
[integrations guide](docs/guides/integrations.md), and the in-depth write-ups in
[`docs/comparisons/`](docs/comparisons).</sub>

## Use cases

| You want to… | Reach for | Example |
|---|---|---|
| Classify and route support tickets into typed labels | typed output | [`01_support_triage.py`](examples/01_support_triage.py) |
| Answer questions over your docs with real citations | hybrid RAG + grounding policy | [`02_document_qa.py`](examples/02_document_qa.py) |
| Review contracts clause-by-clause | end-to-end context app | [`03_contract_review.py`](examples/03_contract_review.py) |
| Extract structured fields from invoices | structured extraction + F1 eval | [`04_invoice_extraction.py`](examples/04_invoice_extraction.py) |
| Build a research agent with bounded budgets | ReAct agent + tools | [`05_research_agent.py`](examples/05_research_agent.py) |
| Automate a CRM agent with approval-gated writes | memory + permissioned tools | [`06_crm_agent.py`](examples/06_crm_agent.py) |
| Ask questions over a codebase | code-aware chunking + import graph | [`07_codebase_qa.py`](examples/07_codebase_qa.py) |
| Analyze spreadsheets with schema awareness | table chunking + quality checks | [`08_spreadsheet_analysis.py`](examples/08_spreadsheet_analysis.py) |
| Gate quality in CI | datasets, gates, baseline diff | [`09_eval_pipeline.py`](examples/09_eval_pipeline.py) |
| Tune prompts/context against an eval suite | optimization + gated promotion | [`10_optimization_run.py`](examples/10_optimization_run.py) |
| Stream answers token-by-token through the full pipeline | `astream` + partial-JSON + compile caches | [`11_streaming_performance.py`](examples/11_streaming_performance.py) |
| Push retrieval quality with the full retrieval toolkit | sparse+late-interaction fusion, HyDE, auto-merge, GraphRAG, connectors | [`12_advanced_rag.py`](examples/12_advanced_rag.py) |
| Personalize an app with governed memory | scoped remember/recall, consolidation, hygiene, memory evals | [`13_memory_personalization.py`](examples/13_memory_personalization.py) |
| Evaluate, test, and observe without a platform | quality metrics, synthetic data, red-teaming, experiments, prompt registry, sessions + trace viewer | [`14_evaluation_observability.py`](examples/14_evaluation_observability.py) |
| Run a multi-agent team with roles and delegation | crews + shared blackboard + budget guarantees | [`15_multi_agent_crew.py`](examples/15_multi_agent_crew.py) |
| Build an interruptible, auditable, resumable process | durable graphs + human gates + time-travel | [`16_durable_graph.py`](examples/16_durable_graph.py) |
| Guarantee output shape and guard every generation | signatures, constrained decoding, streaming validation, rails, self-correction, schema routing | [`17_reliable_structured_output.py`](examples/17_reliable_structured_output.py) |
| Improve the app from its own production traffic | the closed loop: traces→dataset→eval→optimize→promote, auto-memory, retrieval feedback, Pareto, learned budgets | [`18_closed_loop.py`](examples/18_closed_loop.py) |
| Reuse LangChain/LlamaIndex assets and any OpenAI-compatible model | framework interop + provider/vector-store breadth | [`19_framework_interop.py`](examples/19_framework_interop.py) |
| Configure an app for a domain in one line | opt-in domain packs (support/engineering/finance/legal) | [`20_domain_pack.py`](examples/20_domain_pack.py) |
| Govern PII, injection, access, and audit integrity | deterministic security primitives + tamper-evident audit | [`21_security_governance.py`](examples/21_security_governance.py) |
| Use MCP servers as tools/resources, or expose your app as one | MCP client + server | [`22_mcp_tools_and_resources.py`](examples/22_mcp_tools_and_resources.py) |
| Expose a crew over A2A and delegate across vendor agents | A2A agent card + task lifecycle + remote delegate | [`23_a2a_delegation.py`](examples/23_a2a_delegation.py) |
| Drop in portable `SKILL.md` knowledge with budgeting | Agent Skills + progressive disclosure | [`24_agent_skills.py`](examples/24_agent_skills.py) |
| Control reasoning effort across providers with honest cost | unified reasoning control + Responses API | [`25_reasoning_control.py`](examples/25_reasoning_control.py) |
| Score agents over their trajectory and live traffic | trajectory & tool-use metrics, multi-turn simulator, online eval + drift, Cohen's-κ annotation, A/B + metric-as-guardrail | [`26_agentic_eval.py`](examples/26_agentic_eval.py) |
| Survive outages and account for every dollar at scale | batch execution, circuit breaking + failover, key pooling, model cascades, cost attribution + budgets, prompt caching, sharded indexing | [`27_cost_and_reliability.py`](examples/27_cost_and_reliability.py) |
| Optimize prompts reflectively and distill traces into a cheaper model | GEPA/MIPRO reflective optimizer, distillation flywheel, learned compression, optimizer-judge calibration | [`28_reflective_optimization.py`](examples/28_reflective_optimization.py) |
| Shrink embeddings, retrieve across text+image, and add stores | Matryoshka (MRL) truncation, contextual & multimodal embedders, new vector stores, layout-aware extraction, voice/realtime | [`29_multimodal_retrieval.py`](examples/29_multimodal_retrieval.py) |
| Generate compliance evidence and satisfy a data-erasure request | model/system cards, OWASP/NIST/MITRE mapping, AI-BOM, lineage + erasure, residency, multilingual PII | [`30_governance_compliance.py`](examples/30_governance_compliance.py) |

## More examples

All thirty examples in [`examples/`](examples) run **fully offline** with no API keys. Point them
at a real model with environment variables:

```bash
export VINCIO_PROVIDER=openai VINCIO_MODEL=gpt-5.2-mini OPENAI_API_KEY=sk-...
cd examples && python 02_document_qa.py
```

## Command line

```bash
vincio init my-project --template rag  # scaffold config + app + golden set (minimal|rag|agent|eval)
vincio config schema --output vincio.schema.json  # typed JSON Schema for editor completion
vincio packs list                # opt-in domain packs (support/engineering/finance/legal)
vincio tui                       # interactive inspector for runs, traces, and memory
vincio run app.py --input "..."  # run an app
vincio eval run golden.jsonl     # run an eval suite (with CI gates and baseline compare)
vincio eval dataset golden.jsonl --min-feedback 0.5  # curate traces into a dataset
vincio prompt lint prompts/      # lint prompt specs
vincio prompt push prompts/support.yaml --tag production  # version a prompt
vincio trace view trace_123      # TUI trace tree with scores + feedback
vincio trace export trace_123    # self-contained static HTML (also --session)
vincio trace diff a b --html diff.html  # visual side-by-side diff
vincio trace sessions            # list sessions with aggregates
vincio trace feedback trace_123 --score 1.0
vincio optimize run --target groundedness
vincio optimize reflective --app app.py --dataset golden.jsonl  # GEPA-style reflective optimization
vincio loop run --app app.py --min-feedback 0.5 --gate groundedness=">= 0.8"  # one closed-loop cycle
vincio distill --traces-dir .vincio/traces --output train.jsonl  # grounded fine-tuning JSONL
vincio index build ./docs        # build a retrieval index
vincio memory inspect --user u1  # inspect a user's memory
vincio memory recall "answer style" --user u1  # scored hybrid recall
vincio audit verify              # verify the audit-log hash chain offline
vincio mcp tools --command "python server.py"  # inspect an MCP server's tools
vincio mcp serve app.py          # expose an app as an MCP server (stdio)
```

A FastAPI server (API-key + JWT auth, real-token SSE streaming) is available via
`from vincio.server import create_app` — see [`docs/reference/api.md`](docs/reference/api.md).

## Architecture

```text
                         ┌──────────────────────────────────────────────┐
   user input  ─────────▶│  Input engine   normalize · classify · scope  │
                         └───────────────┬──────────────────────────────┘
                                         ▼
        ┌──────────────┐        ┌────────────────┐        ┌──────────────┐
        │   Memory     │───────▶│    CONTEXT     │◀───────│  Retrieval   │
        │  L0…L5       │        │   COMPILER     │        │  hybrid RAG  │
        └──────────────┘        │ score·dedupe·  │        └──────────────┘
        ┌──────────────┐        │ conflict·      │        ┌──────────────┐
        │    Tools     │───────▶│ compress·budget│◀───────│   Prompt     │
        │ permissioned │        └───────┬────────┘        │  compiler    │
        └──────────────┘                ▼                 └──────────────┘
                              ┌────────────────────┐
                              │   Model execution  │   provider-neutral
                              └─────────┬──────────┘
                                        ▼
                    ┌─────────────────────────────────────────┐
                    │ Output validation · Evals · Security ·   │
                    │ Trace + cost · Memory write-back         │
                    └─────────────────────────────────────────┘
```

See [`AGENTS.md`](AGENTS.md) for the package layout and [`docs/concepts/`](docs/concepts) for a tour
of each engine.

## Roadmap

Every subsystem above is implemented, tested offline, documented, and demonstrated by a runnable
example. The public API is frozen under [Semantic Versioning](https://semver.org/spec/v2.0.0.html)
with a mechanical [deprecation policy](docs/reference/stability.md); performance and quality targets
are [published as SLOs](docs/reference/slo.md) and gated by VincioBench; the
[threat model](docs/security/threat-model.md) is documented with offline audit-chain verification and
a resource-limited tool sandbox; and releases ship a CycloneDX SBOM with SLSA provenance attestations.

New capabilities are added without breaking working code: each one sits behind a new entry point or an
opt-in extra, and unproven surface is marked [`@experimental`](docs/reference/stability.md). Vincio
adopts the ecosystem's standards — the MCP, A2A, and Agent Skills protocols, and the OWASP LLM 2025 /
OWASP Agentic / NIST AI RMF / MITRE ATLAS governance frameworks — *in your process*; it never becomes
a hosted service to do so.

**1.7 makes the spine's promises literally true:** the advertised `Budget` is a hard cap, the compiler
scores and selects context by embedding cosine (with MMR and value-level contradiction) when you opt
in, the streaming and non-streaming run paths are unified behind one cooperative-cancellation
epilogue, and a data-driven `ModelRegistry` (capabilities, pricing, lifecycle) replaces substring
guesswork — every promotion now gated on statistical significance, with a trace-replay executor for
reproducible behavioral regression. **1.8 turns that registry into a rotation-and-regression
discipline:** capability guards refuse a substitution that can't serve the request, a `SwapGate`
replays golden traces and diffs quality/cost/latency/behavior with significance to PASS/FAIL a model
swap, a shadow provider and a capped canary qualify a candidate on live traffic with automatic
rollback, and a lifecycle watcher proposes migrations off deprecated models. The arc from here to
**3.0** continues additively: making **documents and images flow *out*** as cited, governed,
eval-gated artifacts (not just in), and closing the improvement loop into **continual, online,
safely-rolled-back self-improvement** — with 2.0 and 3.0 the only breaking windows.

See **[ROADMAP.md](ROADMAP.md)** for what ships today, what's planned, and what's intentionally out of
scope.

Vincio is, and stays, a **library**. The building blocks for production operation (audit chain,
retention, tenant isolation, RBAC/ABAC, a server) ship in the package for you to deploy on your own
infrastructure. Hosted services and managed control planes are not part of this project.

## Documentation

- **[Getting started](docs/getting-started.md)** — install, your first app, offline development
- **Concepts** — [context packets](docs/concepts/context-packets.md) ·
  [prompt compiler](docs/concepts/prompt-compiler.md) · [memory](docs/concepts/memory.md) ·
  [retrieval](docs/concepts/retrieval.md) · [agents & workflows](docs/concepts/agents.md) ·
  [evaluation](docs/concepts/evals.md) · [observability](docs/concepts/observability.md)
- **Guides** — [build a RAG app](docs/guides/build-rag-app.md) ·
  [connect data sources](docs/guides/connectors.md) ·
  [structured output](docs/guides/structured-output.md) ·
  [reliability & guardrails](docs/guides/reliability-guardrails.md) ·
  [add tools](docs/guides/add-tools.md) ·
  [orchestrate multi-agent systems](docs/guides/orchestrate-agents.md) ·
  [run evals](docs/guides/run-evals.md) · [test LLM apps](docs/guides/test-llm-apps.md) ·
  [optimize](docs/guides/optimize-context.md) · [close the loop](docs/guides/close-the-loop.md) ·
  [performance & streaming](docs/guides/performance.md) ·
  [cost, reliability & scale](docs/guides/cost-and-reliability.md) ·
  [integrations](docs/guides/integrations.md)
- **Agentic evaluation & continuous quality** —
  [trajectory metrics, simulator, online eval, drift & annotation](docs/guides/agentic-eval.md)
- **Protocols & interoperability** — [MCP client + server](docs/guides/mcp.md) ·
  [A2A agent-to-agent](docs/guides/a2a.md) · [Agent Skills](docs/guides/agent-skills.md) ·
  [reasoning control & Responses API](docs/guides/reasoning.md) ·
  [voice & realtime](docs/guides/realtime.md)
- **Migrating** — coming from [LangChain](docs/guides/migrate-from-langchain.md) ·
  [LlamaIndex](docs/guides/migrate-from-llamaindex.md) ·
  [Ragas](docs/guides/migrate-from-ragas.md) · [Mem0](docs/guides/migrate-from-mem0.md)
- **Reference** — [API](docs/reference/api.md) · [CLI](docs/reference/cli.md) ·
  [config](docs/reference/config.md) · [API stability & deprecation policy](docs/reference/stability.md) ·
  [performance & quality SLOs](docs/reference/slo.md)
- **Security & governance** — [threat model](docs/security/threat-model.md) ·
  [security policy](SECURITY.md) · [reliability & guardrails guide](docs/guides/reliability-guardrails.md) ·
  [governance & compliance](docs/guides/governance.md)
- **Comparisons** — [LangChain](docs/comparisons/langchain.md) ·
  [LlamaIndex](docs/comparisons/llamaindex.md) · [RAGatouille](docs/comparisons/ragatouille.md) ·
  [Mem0](docs/comparisons/mem0.md) · [CrewAI](docs/comparisons/crewai.md) ·
  [OpenAI Agents SDK](docs/comparisons/openai-agents-sdk.md) · [DSPy](docs/comparisons/dspy.md) ·
  [Pydantic AI](docs/comparisons/pydantic-ai.md) · [Guardrails AI](docs/comparisons/guardrails.md) ·
  [NeMo Guardrails](docs/comparisons/nemo-guardrails.md) · [Ragas](docs/comparisons/ragas.md) ·
  [LiteLLM / gateways](docs/comparisons/litellm.md)

## Contributing

Contributions are welcome. The test suite runs fully offline and must stay green:

```bash
pip install -e ".[dev]"
python -m pytest tests/ -q     # 740 tests, no network or API keys required
ruff check vincio/ tests/
```

See [`AGENTS.md`](AGENTS.md) for the codebase layout and engineering conventions.

## License

[Apache License 2.0](LICENSE) © Vincio Contributors.
