pyveil 0.2.2__tar.gz → 0.2.3__tar.gz

{pyveil-0.2.2 → pyveil-0.2.3}/AGENTS.md

@@ -19,6 +19,7 @@ Do not turn pyveil into a general DLP suite, gateway, Presidio clone, or prompt-
 - Run CLI demo: `python3 -m pyveil demo`
 - Run CLI locally: `PYVEIL_SECRET=dev-secret python3 -m pyveil redact <file>`
 - Run detector evaluation: `python3 evaluation/evaluate.py --check`
+- Run the Ollama boundary without a model call: `PYVEIL_SECRET=dev-secret python3 -m pyveil.integrations.ollama --dry-run`
 
 ## Architecture
 
@@ -72,6 +73,7 @@ This repository includes files meant for coding agents and LLM readers:
 - [docs/evaluation.html](docs/evaluation.html)
 - [docs/guides/](docs/guides/)
 - [docs/integrations/mcp.md](docs/integrations/mcp.md)
+- [docs/integrations/ollama.md](docs/integrations/ollama.md)
 - [docs/integrations/logging.md](docs/integrations/logging.md)
 - [docs/integrations/tracing.md](docs/integrations/tracing.md)
 
@@ -88,6 +90,7 @@ Examples:
 - [examples/fastapi_middleware.py](examples/fastapi_middleware.py)
 - [examples/litellm_proxy_filter.py](examples/litellm_proxy_filter.py)
 - [examples/azure_openai.py](examples/azure_openai.py)
+- [examples/ollama.py](examples/ollama.py)
 
 ## Development
 

{pyveil-0.2.2 → pyveil-0.2.3}/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## 0.2.3 - 2026-07-16
+
+Local Ollama boundary release.
+
+### Added
+
+- Installable Ollama chat integration with env, `.env`, and YAML configuration,
+  a dry-run boundary, response metrics, and configurable context/keep-alive limits.
+- A tested `qwen3.5:4b` local-model recipe sized for a 16GB Apple silicon Mac.
+
+### Changed
+
+- Added an optional `ollama` dependency group while keeping the core package
+  dependency-free.
+- Expanded the README, Cookbook, web Manual, and agent-facing navigation with
+  a complete local-model request path and measured memory/latency trade-offs.
+
 ## 0.2.2 - 2026-07-15
 
 Azure OpenAI adoption release.

{pyveil-0.2.2 → pyveil-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: pyveil
-Version: 0.2.2
+Version: 0.2.3
 Summary: Local PII and secret redaction for Python LLM apps, AI agents, tool calls, and MCP
 Author-email: Hyeon Sang Jeon <wingnut0310@gmail.com>
 Maintainer-email: Hyeon Sang Jeon <wingnut0310@gmail.com>
@@ -39,6 +39,10 @@ Provides-Extra: azure-openai
 Requires-Dist: openai>=1.66.0; extra == "azure-openai"
 Requires-Dist: python-dotenv>=1.0.0; extra == "azure-openai"
 Requires-Dist: PyYAML>=6.0.0; extra == "azure-openai"
+Provides-Extra: ollama
+Requires-Dist: ollama<1.0.0,>=0.6.2; extra == "ollama"
+Requires-Dist: python-dotenv>=1.0.0; extra == "ollama"
+Requires-Dist: PyYAML>=6.0.0; extra == "ollama"
 Provides-Extra: dev
 Requires-Dist: build>=1.2.0; extra == "dev"
 Requires-Dist: pytest>=7.0.0; extra == "dev"
@@ -47,6 +51,7 @@ Requires-Dist: black>=23.0.0; extra == "dev"
 Requires-Dist: ruff>=0.1.0; extra == "dev"
 Requires-Dist: twine>=5.0.0; extra == "dev"
 Requires-Dist: mypy>=1.0.0; extra == "dev"
+Requires-Dist: ollama<1.0.0,>=0.6.2; extra == "dev"
 Requires-Dist: openai>=1.66.0; extra == "dev"
 Requires-Dist: python-dotenv>=1.0.0; extra == "dev"
 Requires-Dist: PyYAML>=6.0.0; extra == "dev"
@@ -140,6 +145,70 @@ response = call_llm(safe.data)  # Your provider SDK call
 The provider receives the same list and dictionary shape, with sensitive values
 replaced before serialization or transmission.
 
+## Ollama: Local End To End
+
+Run a local model behind the same redaction boundary. The optional integration
+uses Ollama's official Python client and defaults to `qwen3.5:4b`, a Q4_K_M
+4.7B model that fits comfortably on a 16GB Apple silicon Mac with a 4K context:
+
+```bash
+pip install "pyveil[ollama]"
+ollama pull qwen3.5:4b
+```
+
+```python
+from pyveil.integrations.ollama import ask_ollama, load_settings
+
+settings = load_settings()  # OLLAMA_* + PYVEIL_* environment variables
+result = ask_ollama(
+    "Write a follow-up for alice@example.com or 010-1234-5678.",
+    settings,
+)
+
+print(result.redacted_input)  # The exact prompt sent to Ollama
+print(result.output_text)     # The local model response
+```
+
+Prove the boundary without loading a model:
+
+```bash
+PYVEIL_SECRET=docs-demo-secret \
+  python -m pyveil.integrations.ollama --dry-run
+```
+
+```text
+mode: dry-run
+model: qwen3.5:4b
+host: http://127.0.0.1:11434
+sent-to-ollama: Write a one-sentence support follow-up for [EMAIL:71c6727a7fa2] or [PHONE:b4b889df07ce].
+findings: EMAIL=1, PHONE=1
+ollama-response: skipped (--dry-run)
+```
+
+For a live local call, set `PYVEIL_SECRET` and run the module. Configuration
+priority is process environment, `.env`, YAML, then defaults:
+
+```bash
+PYVEIL_SECRET=a-long-random-hmac-secret \
+  python -m pyveil.integrations.ollama
+
+python -m pyveil.integrations.ollama \
+  --config examples/ollama.example.yaml --env-file .env
+```
+
+The checked-in [`.env` template](examples/ollama.env.example) and
+[YAML template](examples/ollama.example.yaml) expose model, host, context,
+output length, temperature, timeout, and keep-alive. pyveil caps the default
+context at 4096 and uses `keep_alive=0`, so one-shot calls release model memory;
+set `OLLAMA_KEEP_ALIVE=5m` for faster repeated calls.
+
+Observed on this project's M1 Mac mini with 16GB memory and Ollama 0.31.2:
+`qwen3.5:4b` used about 3.2GB at a 4096-token context, a cold request took
+8.1 seconds, and a warm request took 1.3 seconds. These are local measurements,
+not portable performance guarantees. See the
+[Ollama integration guide](docs/integrations/ollama.md) for the full setup and
+memory trade-offs.
+
 ## Azure OpenAI: End To End
 
 Install the optional Azure example dependencies, then load configuration from
@@ -226,7 +295,7 @@ evaluator:
 python evaluation/evaluate.py --check
 ```
 
-For corpus v1, pyveil 0.2.2 matches all 36 expected findings across 39 cases
+For corpus v1, pyveil 0.2.3 matches all 36 expected findings across 39 cases
 (33 positive, 6 negative), with no corpus false positives, false negatives,
 labeled-value leaks, or non-empty `Finding.raw` values.
 

{pyveil-0.2.2 → pyveil-0.2.3}/README.md

@@ -81,6 +81,70 @@ response = call_llm(safe.data)  # Your provider SDK call
 The provider receives the same list and dictionary shape, with sensitive values
 replaced before serialization or transmission.
 
+## Ollama: Local End To End
+
+Run a local model behind the same redaction boundary. The optional integration
+uses Ollama's official Python client and defaults to `qwen3.5:4b`, a Q4_K_M
+4.7B model that fits comfortably on a 16GB Apple silicon Mac with a 4K context:
+
+```bash
+pip install "pyveil[ollama]"
+ollama pull qwen3.5:4b
+```
+
+```python
+from pyveil.integrations.ollama import ask_ollama, load_settings
+
+settings = load_settings()  # OLLAMA_* + PYVEIL_* environment variables
+result = ask_ollama(
+    "Write a follow-up for alice@example.com or 010-1234-5678.",
+    settings,
+)
+
+print(result.redacted_input)  # The exact prompt sent to Ollama
+print(result.output_text)     # The local model response
+```
+
+Prove the boundary without loading a model:
+
+```bash
+PYVEIL_SECRET=docs-demo-secret \
+  python -m pyveil.integrations.ollama --dry-run
+```
+
+```text
+mode: dry-run
+model: qwen3.5:4b
+host: http://127.0.0.1:11434
+sent-to-ollama: Write a one-sentence support follow-up for [EMAIL:71c6727a7fa2] or [PHONE:b4b889df07ce].
+findings: EMAIL=1, PHONE=1
+ollama-response: skipped (--dry-run)
+```
+
+For a live local call, set `PYVEIL_SECRET` and run the module. Configuration
+priority is process environment, `.env`, YAML, then defaults:
+
+```bash
+PYVEIL_SECRET=a-long-random-hmac-secret \
+  python -m pyveil.integrations.ollama
+
+python -m pyveil.integrations.ollama \
+  --config examples/ollama.example.yaml --env-file .env
+```
+
+The checked-in [`.env` template](examples/ollama.env.example) and
+[YAML template](examples/ollama.example.yaml) expose model, host, context,
+output length, temperature, timeout, and keep-alive. pyveil caps the default
+context at 4096 and uses `keep_alive=0`, so one-shot calls release model memory;
+set `OLLAMA_KEEP_ALIVE=5m` for faster repeated calls.
+
+Observed on this project's M1 Mac mini with 16GB memory and Ollama 0.31.2:
+`qwen3.5:4b` used about 3.2GB at a 4096-token context, a cold request took
+8.1 seconds, and a warm request took 1.3 seconds. These are local measurements,
+not portable performance guarantees. See the
+[Ollama integration guide](docs/integrations/ollama.md) for the full setup and
+memory trade-offs.
+
 ## Azure OpenAI: End To End
 
 Install the optional Azure example dependencies, then load configuration from
@@ -167,7 +231,7 @@ evaluator:
 python evaluation/evaluate.py --check
 ```
 
-For corpus v1, pyveil 0.2.2 matches all 36 expected findings across 39 cases
+For corpus v1, pyveil 0.2.3 matches all 36 expected findings across 39 cases
 (33 positive, 6 negative), with no corpus false positives, false negatives,
 labeled-value leaks, or non-empty `Finding.raw` values.
 

{pyveil-0.2.2 → pyveil-0.2.3}/docs/README.md

@@ -12,6 +12,7 @@ Start with:
 - [Threat model](threat-model.md)
 - [FAQ](faq.md)
 - [Cookbook](cookbook.md)
+- [Ollama local-model integration](integrations/ollama.md)
 - [Azure OpenAI env/YAML integration](../pyveil/integrations/azure_openai.py)
 - [Redaction reference](redaction-reference.md)
 - [Known limitations](known-limitations.md)

{pyveil-0.2.2 → pyveil-0.2.3}/docs/cookbook.md

@@ -48,7 +48,92 @@ response = call_with_redaction(
 print(response)
 ```
 
-## 2. Azure OpenAI With Environment Or YAML Configuration
+## 2. Ollama With A Memory-Aware Local Model
+
+Install the optional dependencies and pull the recommended 16GB-Mac model:
+
+```bash
+pip install "pyveil[ollama]"
+ollama pull qwen3.5:4b
+```
+
+The helper redacts before `Client.chat(...)` and returns the exact local-model
+input, response text, and available timing/token metrics:
+
+```python
+from pyveil.integrations.ollama import ask_ollama, load_settings
+
+
+def answer_with_ollama(prompt: str) -> str:
+    settings = load_settings()
+    result = ask_ollama(prompt, settings)
+    print("sent-to-ollama:", result.redacted_input)
+    print("total-ms:", result.total_duration_ms)
+    return result.output_text or ""
+
+
+answer = answer_with_ollama(
+    "Write a follow-up for alice@example.com or 010-1234-5678."
+)
+```
+
+Under the hood, the boundary stays deliberately small:
+
+```python
+from ollama import Client
+from pyveil import Channel, Veil
+
+client = Client(host=settings.host, timeout=settings.timeout_seconds)
+veil = Veil.high(
+    secret=settings.pyveil_secret,
+    scope=settings.pyveil_scope,
+)
+
+safe = veil.redact_text(prompt, channel=Channel.PROMPT_INPUT)
+response = client.chat(
+    model=settings.model,
+    messages=[{"role": "user", "content": safe.text}],
+    stream=False,
+    think=False,
+    options={"num_ctx": 4096, "num_predict": 128, "temperature": 0.2},
+    keep_alive="0",
+)
+```
+
+Try the non-model and live paths:
+
+```bash
+export PYVEIL_SECRET="a-long-random-hmac-secret"
+
+python -m pyveil.integrations.ollama --dry-run
+python -m pyveil.integrations.ollama
+python -m pyveil.integrations.ollama \
+  --config examples/ollama.example.yaml --env-file .env
+```
+
+Configuration priority is process environment, `.env`, YAML, then defaults.
+The default `qwen3.5:4b` configuration uses a 4096-token context and
+`keep_alive=0` to release memory after each response. Use
+`OLLAMA_KEEP_ALIVE=5m` when repeated-call latency matters more than reclaiming
+roughly 3.2GB immediately on a 16GB Apple silicon Mac.
+
+Dry-run output from the synthetic default prompt:
+
+```text
+mode: dry-run
+model: qwen3.5:4b
+host: http://127.0.0.1:11434
+sent-to-ollama: Write a one-sentence support follow-up for [EMAIL:71c6727a7fa2] or [PHONE:b4b889df07ce].
+findings: EMAIL=1, PHONE=1
+ollama-response: skipped (--dry-run)
+```
+
+See [`pyveil/integrations/ollama.py`](../pyveil/integrations/ollama.py),
+[`examples/ollama.env.example`](../examples/ollama.env.example),
+[`examples/ollama.example.yaml`](../examples/ollama.example.yaml), and the
+[full Ollama integration guide](integrations/ollama.md).
+
+## 3. Azure OpenAI With Environment Or YAML Configuration
 
 Install the optional dependencies:
 
@@ -137,7 +222,7 @@ See [`pyveil/integrations/azure_openai.py`](../pyveil/integrations/azure_openai.
 [`examples/azure_openai.env.example`](../examples/azure_openai.env.example), and
 [`examples/azure_openai.example.yaml`](../examples/azure_openai.example.yaml).
 
-## 3. Block Credentials Before Model-Controlled Tool Calls
+## 4. Block Credentials Before Model-Controlled Tool Calls
 
 The default policy blocks credential-like material in `tool.call.arguments`.
 
@@ -160,7 +245,7 @@ else:
     run_tool(**safe_args)
 ```
 
-## 4. Redact Tool Results Before Returning Them To The Model
+## 5. Redact Tool Results Before Returning Them To The Model
 
 Tool results often contain customer records, logs, URLs, headers, or hidden tokens.
 
@@ -177,7 +262,7 @@ raw_result = {
 safe_result = veil.redact_data(raw_result, channel=Channel.TOOL_CALL_RESULT).data
 ```
 
-## 5. Redact MCP Resource Content
+## 6. Redact MCP Resource Content
 
 MCP resources can expose files, database rows, logs, or API payloads. Redact before the resource content enters an agent context.
 
@@ -192,7 +277,7 @@ def read_resource_for_agent(resource_id: str) -> object:
     return veil.redact_data(raw_resource, channel=Channel.MCP_RESOURCE_CONTENT).data
 ```
 
-## 6. Redact Memory Before Embedding Or Persistence
+## 7. Redact Memory Before Embedding Or Persistence
 
 Memory stores create long-lived recall. Redact before embedding and before persistence.
 
@@ -208,7 +293,7 @@ def write_memory(text: str) -> None:
     memory_store.save(text=safe.text, embedding=embedding)
 ```
 
-## 7. Redact Logs Before Export
+## 8. Redact Logs Before Export
 
 Use the built-in logging integration for application logs.
 
@@ -224,7 +309,7 @@ logger.addFilter(PyVeilLogFilter(Veil.high(secret=b"log-secret", scope="prod/log
 logger.warning("failed request for alice@example.com")
 ```
 
-## 8. CLI Preflight In CI Or Local Scripts
+## 9. CLI Preflight In CI Or Local Scripts
 
 Use `pyveil scan` to count findings and `pyveil redact` to create a redacted artifact.
 
@@ -239,7 +324,7 @@ pyveil redact prompt.txt --channel prompt.input > prompt.safe.txt
 pyveil redact request.json --channel tool.call.result --format json > request.safe.json
 ```
 
-## 9. Redact Known Names And Domain Identifiers
+## 10. Redact Known Names And Domain Identifiers
 
 The dependency-free core does not guess unknown names. Add values that your
 application already knows are sensitive, plus narrow identifiers from your own

{pyveil-0.2.2 → pyveil-0.2.3}/docs/evaluation.html

@@ -47,7 +47,7 @@
       <p class="eyebrow">Evidence, not promises</p>
       <h1>A reproducible detector <strong>regression evaluation</strong>.</h1>
       <p class="lede">The public corpus tests documented supported shapes, negative examples, labeled-value removal, and the no-raw-findings contract on every change.</p>
-      <p class="meta">Corpus v1 &middot; pyveil 0.2.2 &middot; Last run July 15, 2026</p>
+      <p class="meta">Corpus v1 &middot; pyveil 0.2.3 &middot; Last run July 16, 2026</p>
     </div>
   </div>
 

{pyveil-0.2.2 → pyveil-0.2.3}/docs/index.html

@@ -27,7 +27,7 @@
     "name": "pyveil",
     "applicationCategory": "DeveloperApplication",
     "operatingSystem": "Any",
-    "softwareVersion": "0.2.2",
+    "softwareVersion": "0.2.3",
     "programmingLanguage": "Python",
     "description": "Local PII and secret redaction for Python LLM apps and AI agents.",
     "url": "https://hyeonsangjeon.github.io/pyveil/",
@@ -386,6 +386,11 @@ API key [API_KEY:38ded98a17e7]</code></pre>
     <section class="section">
       <h2>Recipes people actually need</h2>
       <div class="grid">
+        <a class="recipe" href="manual.html#ollama">
+          <h3>Ollama, local and memory-aware</h3>
+          <p>Run a tested 4B model on a 16GB Mac and inspect the exact redacted prompt before local inference.</p>
+          <span class="more">Read &rarr;</span>
+        </a>
         <a class="recipe" href="manual.html#azure-openai">
           <h3>Azure OpenAI, end to end</h3>
           <p>Load env or YAML, inspect the exact redacted prompt, then call the Azure v1 Responses API.</p>

pyveil-0.2.3/docs/integrations/ollama.md

@@ -0,0 +1,134 @@
+# Ollama Integration
+
+Use `pyveil.integrations.ollama` when a prompt should be redacted immediately
+before it enters a local Ollama model.
+
+```text
+raw prompt -> pyveil HIGH redaction -> Ollama Client.chat -> local model
+```
+
+## Install
+
+Install the optional official client and configuration dependencies, then pull
+a model:
+
+```bash
+pip install "pyveil[ollama]"
+ollama pull qwen3.5:4b
+```
+
+The integration does not pull a model automatically. Model downloads are large
+and should remain an explicit operator decision.
+
+## Why `qwen3.5:4b` For A 16GB Mac
+
+The development machine used for this integration is an M1 Mac mini with 16GB
+of unified memory. The selected `qwen3.5:4b` tag is a Q4_K_M model with about
+4.7B parameters and a 3.4GB download. The integration limits context to 4096
+tokens even though the model supports more, because KV cache grows with context.
+
+Observed locally with Ollama 0.31.2:
+
+| Check | Observed result |
+| --- | --- |
+| Loaded model size from `ollama ps` | 3.2GB |
+| Processor | 100% GPU |
+| Context | 4096 tokens |
+| Cold live request | 8.1s total, 5.2s model load |
+| Warm live request | 1.3s total, 0.25s load |
+| Memory free indicator while resident | 26% |
+| Memory free indicator after unload | 57% |
+
+These values are a single-machine smoke test, not a benchmark. Background
+applications, prompt length, model updates, and Ollama versions will change
+them.
+
+## Dry Run
+
+Dry-run proves what would cross the client boundary without loading the model:
+
+```bash
+PYVEIL_SECRET=docs-demo-secret \
+  python -m pyveil.integrations.ollama --dry-run
+```
+
+```text
+mode: dry-run
+model: qwen3.5:4b
+host: http://127.0.0.1:11434
+sent-to-ollama: Write a one-sentence support follow-up for [EMAIL:71c6727a7fa2] or [PHONE:b4b889df07ce].
+findings: EMAIL=1, PHONE=1
+ollama-response: skipped (--dry-run)
+```
+
+## Live Call
+
+```bash
+export PYVEIL_SECRET="a-long-random-hmac-secret"
+python -m pyveil.integrations.ollama
+```
+
+Or call the reusable API:
+
+```python
+from pyveil.integrations.ollama import ask_ollama, load_settings
+
+settings = load_settings()
+result = ask_ollama(
+    "Write a support reply for alice@example.com or 010-1234-5678.",
+    settings,
+)
+
+print(result.redacted_input)
+print(result.output_text)
+print(result.total_duration_ms, result.prompt_eval_count, result.eval_count)
+```
+
+`redacted_input` is the exact value passed to `Client.chat`. Findings and
+exceptions do not retain the original email or phone number.
+
+## Configuration
+
+Priority is process environment, `.env`, YAML, then defaults.
+
+| Environment | YAML key | Default | Purpose |
+| --- | --- | --- | --- |
+| `OLLAMA_HOST` | `ollama.host` | `http://127.0.0.1:11434` | Ollama server URL |
+| `OLLAMA_MODEL` | `ollama.model` | `qwen3.5:4b` | Local model tag |
+| `OLLAMA_NUM_CTX` | `ollama.num_ctx` | `4096` | Context and KV-cache limit |
+| `OLLAMA_NUM_PREDICT` | `ollama.num_predict` | `128` | Maximum output tokens |
+| `OLLAMA_TEMPERATURE` | `ollama.temperature` | `0.2` | Sampling temperature |
+| `OLLAMA_KEEP_ALIVE` | `ollama.keep_alive` | `0` | How long the model stays resident |
+| `OLLAMA_TIMEOUT` | `ollama.timeout_seconds` | `120` | HTTP timeout in seconds |
+| `PYVEIL_SECRET` | `pyveil.secret_env` | required | HMAC secret environment variable |
+| `PYVEIL_SCOPE` | `pyveil.scope` | `ollama/example` | Placeholder namespace |
+
+```bash
+python -m pyveil.integrations.ollama \
+  --config examples/ollama.example.yaml \
+  --env-file .env
+```
+
+The pyveil secret cannot be stored directly in YAML. The loader rejects a
+`pyveil.secret` field; YAML may only name the environment variable that holds
+it.
+
+## Memory And Latency
+
+- Keep the default `OLLAMA_KEEP_ALIVE=0` for one-shot jobs or a memory-busy
+  16GB machine. Ollama unloads the model after the response.
+- Set `OLLAMA_KEEP_ALIVE=5m` for repeated agent turns. Warm requests are much
+  faster, but the model remains resident during that window.
+- Keep `OLLAMA_NUM_CTX=4096` unless the application proves it needs more.
+- Inspect current residency with `ollama ps`; unload manually with
+  `ollama stop qwen3.5:4b`.
+
+## Boundary Notes
+
+- `127.0.0.1` keeps the redacted request on the machine. A remote
+  `OLLAMA_HOST` creates a network boundary even though the raw PII was removed.
+- pyveil protects the input boundary. If model output will be logged, stored,
+  or sent onward, redact it again with channel `prompt.output`.
+- Regex and known-value rules do not discover arbitrary names or addresses.
+- The integration disables thinking output and returns only
+  `message.content`; it does not persist or expose a reasoning trace.

{pyveil-0.2.2 → pyveil-0.2.3}/docs/llms.txt

@@ -43,6 +43,7 @@ EMAIL, PHONE, CREDIT_CARD, JWT, AUTH_HEADER, PRIVATE_KEY, API_KEY, URL_QUERY_SEC
 ## Repository documents
 
 - Cookbook: https://github.com/hyeonsangjeon/pyveil/blob/main/docs/cookbook.md
+- Ollama integration: https://github.com/hyeonsangjeon/pyveil/blob/main/docs/integrations/ollama.md
 - Redaction reference: https://github.com/hyeonsangjeon/pyveil/blob/main/docs/redaction-reference.md
 - Threat model: https://github.com/hyeonsangjeon/pyveil/blob/main/docs/threat-model.md
 - Known limitations: https://github.com/hyeonsangjeon/pyveil/blob/main/docs/known-limitations.md