svitch 0.1.5__tar.gz

svitch-0.1.5/.gitignore

@@ -0,0 +1,30 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+dist/
+build/
+.venv/
+venv/
+*.egg
+
+# Node
+node_modules/
+sdk/node/dist/
+
+# Environment
+.env
+.env.local
+.env.*.local
+
+# IDE
+.vscode/
+.idea/
+*.swp
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Logs
+*.log

svitch-0.1.5/PKG-INFO

@@ -0,0 +1,165 @@
+Metadata-Version: 2.5
+Name: svitch
+Version: 0.1.5
+Summary: Privacy-first AI infrastructure — PII detection, redaction, agent audit trails, and compliance reports for teams building AI on sensitive data.
+Project-URL: Homepage, https://svitch.ai
+Project-URL: Documentation, https://svitch.ai/dpdp
+Project-URL: Repository, https://github.com/koushiknarendra/svitch
+Project-URL: Issues, https://github.com/koushiknarendra/svitch/issues
+License: Apache-2.0
+Keywords: aadhaar,agent,audit,compliance,dpdp,gdpr,hipaa,llm,pan,pii,privacy,security
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Security
+Classifier: Topic :: Software Development :: Libraries
+Requires-Python: >=3.10
+Provides-Extra: all
+Requires-Dist: anthropic>=0.25.0; extra == 'all'
+Requires-Dist: langchain-core>=0.2.0; extra == 'all'
+Requires-Dist: openai>=1.0.0; extra == 'all'
+Requires-Dist: opentelemetry-api>=1.0.0; extra == 'all'
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=0.25.0; extra == 'anthropic'
+Provides-Extra: langchain
+Requires-Dist: langchain-core>=0.2.0; extra == 'langchain'
+Provides-Extra: openai
+Requires-Dist: openai>=1.0.0; extra == 'openai'
+Provides-Extra: otel
+Requires-Dist: opentelemetry-api>=1.0.0; extra == 'otel'
+Description-Content-Type: text/markdown
+
+# svitch
+
+DPDP-compliant AI data security for Indian enterprises.  
+Zero dependencies. Runs locally. Works with any LLM provider.
+
+```bash
+pip install svitch
+```
+
+---
+
+## PII Detection and Redaction
+
+Detects and redacts Indian PII (Aadhaar, PAN, UPI, IFSC, mobile, GST, bank accounts) and global PII (email, IP) — entirely locally, no network calls.
+
+```python
+import svitch
+
+# Detect
+entities = svitch.detect("Customer Aadhaar: 2345 6789 0123, PAN: ABCDE1234F")
+# [Entity(type='AADHAAR', value='2345 6789 0123', ...), Entity(type='PAN', ...)]
+
+# Redact (token replacement)
+result = svitch.redact("Call me on 9876543210, UPI: rahul@okicici")
+result.text   # "Call me on [MOBILE_IN], UPI: [UPI_ID]"
+result.count  # 2
+result.clean  # False
+
+# Redact (partial mask)
+result = svitch.redact("Aadhaar: 2345 6789 0123", replacement="mask")
+result.text   # "Aadhaar: XXXX XXXX 0123"
+```
+
+### Wrap any LLM client
+
+PII is redacted from every prompt before it leaves your network, and from every response before it's stored or displayed.
+
+```python
+import svitch, openai
+
+client = svitch.wrap(openai.OpenAI())
+# Use exactly like openai.OpenAI() — PII is handled automatically
+
+response = client.chat.completions.create(
+    model="gpt-4o",
+    messages=[{
+        "role": "user",
+        "content": "Assess loan for Aadhaar 9876 5432 1098, PAN ABCDE1234F"
+        # ↑ redacted to [AADHAAR_IN], [PAN_IN] before reaching OpenAI
+    }]
+)
+```
+
+Anthropic:
+```python
+import svitch, anthropic
+client = svitch.wrap(anthropic.Anthropic())
+```
+
+---
+
+## Agent Audit Trail
+
+Records every agent decision in an immutable, hash-chained audit log.  
+Required by DPDP §8 and the RBI FREE AI Framework.
+
+```python
+from svitch_tracer import SvitchTracer
+
+tracer = SvitchTracer(agent_id="loan-processor-v2")
+
+with tracer.run() as run:
+    run.data_access(
+        source="crm",
+        fields_accessed=["name", "income", "aadhaar"],
+        purpose="loan_processing",
+        data_principal_id="CUST-5821",
+    )
+
+    run.llm_call(
+        provider="openai",
+        model="gpt-4o",
+        prompt="Assess eligibility for [AADHAAR_IN] applicant",   # already redacted
+        response="Eligible. Score: 72/100.",
+        redact_pii=True,
+    )
+
+    run.decision(
+        reason="Score above threshold (70)",
+        outcome="approve",
+        confidence=0.87,
+    )
+
+    # Human-in-the-loop checkpoint — required for high-risk decisions
+    run.human_checkpoint(
+        question="Approve ₹5L loan for this applicant?",
+        approved=True,
+        reviewer_id="anand.k",
+    )
+
+# Verify the hash chain is intact
+valid, err = run.verify()
+assert valid, f"Audit chain broken: {err}"
+```
+
+### Configuration
+
+```bash
+export SVITCH_TRACER_URL=https://agent-tracer.vercel.app  # default (hosted)
+# or point to your self-hosted Agent Tracer
+```
+
+---
+
+## Supported PII types
+
+| Type | Pattern |
+|------|---------|
+| `AADHAAR` | 12-digit, masked and unmasked |
+| `PAN` | `ABCDE1234F` format |
+| `UPI_ID` | `handle@provider` |
+| `IFSC` | 11-character bank code |
+| `MOBILE_IN` | 6xxx–9xxx Indian mobile |
+| `BANK_ACCOUNT` | 9–18 digit account numbers |
+| `GST` | `22AAAAA0000A1Z5` |
+| `EMAIL` | RFC 5322 |
+| `IPV4` | IPv4 addresses |
+
+---
+
+## License
+
+Apache 2.0 — [svitch.ai](https://svitch.ai) · [DPDP Guide](https://svitch.ai/dpdp) · [GitHub](https://github.com/koushiknarendra/svitch)

svitch-0.1.5/README.md

@@ -0,0 +1,133 @@
+# svitch
+
+DPDP-compliant AI data security for Indian enterprises.  
+Zero dependencies. Runs locally. Works with any LLM provider.
+
+```bash
+pip install svitch
+```
+
+---
+
+## PII Detection and Redaction
+
+Detects and redacts Indian PII (Aadhaar, PAN, UPI, IFSC, mobile, GST, bank accounts) and global PII (email, IP) — entirely locally, no network calls.
+
+```python
+import svitch
+
+# Detect
+entities = svitch.detect("Customer Aadhaar: 2345 6789 0123, PAN: ABCDE1234F")
+# [Entity(type='AADHAAR', value='2345 6789 0123', ...), Entity(type='PAN', ...)]
+
+# Redact (token replacement)
+result = svitch.redact("Call me on 9876543210, UPI: rahul@okicici")
+result.text   # "Call me on [MOBILE_IN], UPI: [UPI_ID]"
+result.count  # 2
+result.clean  # False
+
+# Redact (partial mask)
+result = svitch.redact("Aadhaar: 2345 6789 0123", replacement="mask")
+result.text   # "Aadhaar: XXXX XXXX 0123"
+```
+
+### Wrap any LLM client
+
+PII is redacted from every prompt before it leaves your network, and from every response before it's stored or displayed.
+
+```python
+import svitch, openai
+
+client = svitch.wrap(openai.OpenAI())
+# Use exactly like openai.OpenAI() — PII is handled automatically
+
+response = client.chat.completions.create(
+    model="gpt-4o",
+    messages=[{
+        "role": "user",
+        "content": "Assess loan for Aadhaar 9876 5432 1098, PAN ABCDE1234F"
+        # ↑ redacted to [AADHAAR_IN], [PAN_IN] before reaching OpenAI
+    }]
+)
+```
+
+Anthropic:
+```python
+import svitch, anthropic
+client = svitch.wrap(anthropic.Anthropic())
+```
+
+---
+
+## Agent Audit Trail
+
+Records every agent decision in an immutable, hash-chained audit log.  
+Required by DPDP §8 and the RBI FREE AI Framework.
+
+```python
+from svitch_tracer import SvitchTracer
+
+tracer = SvitchTracer(agent_id="loan-processor-v2")
+
+with tracer.run() as run:
+    run.data_access(
+        source="crm",
+        fields_accessed=["name", "income", "aadhaar"],
+        purpose="loan_processing",
+        data_principal_id="CUST-5821",
+    )
+
+    run.llm_call(
+        provider="openai",
+        model="gpt-4o",
+        prompt="Assess eligibility for [AADHAAR_IN] applicant",   # already redacted
+        response="Eligible. Score: 72/100.",
+        redact_pii=True,
+    )
+
+    run.decision(
+        reason="Score above threshold (70)",
+        outcome="approve",
+        confidence=0.87,
+    )
+
+    # Human-in-the-loop checkpoint — required for high-risk decisions
+    run.human_checkpoint(
+        question="Approve ₹5L loan for this applicant?",
+        approved=True,
+        reviewer_id="anand.k",
+    )
+
+# Verify the hash chain is intact
+valid, err = run.verify()
+assert valid, f"Audit chain broken: {err}"
+```
+
+### Configuration
+
+```bash
+export SVITCH_TRACER_URL=https://agent-tracer.vercel.app  # default (hosted)
+# or point to your self-hosted Agent Tracer
+```
+
+---
+
+## Supported PII types
+
+| Type | Pattern |
+|------|---------|
+| `AADHAAR` | 12-digit, masked and unmasked |
+| `PAN` | `ABCDE1234F` format |
+| `UPI_ID` | `handle@provider` |
+| `IFSC` | 11-character bank code |
+| `MOBILE_IN` | 6xxx–9xxx Indian mobile |
+| `BANK_ACCOUNT` | 9–18 digit account numbers |
+| `GST` | `22AAAAA0000A1Z5` |
+| `EMAIL` | RFC 5322 |
+| `IPV4` | IPv4 addresses |
+
+---
+
+## License
+
+Apache 2.0 — [svitch.ai](https://svitch.ai) · [DPDP Guide](https://svitch.ai/dpdp) · [GitHub](https://github.com/koushiknarendra/svitch)

svitch-0.1.5/pyproject.toml

@@ -0,0 +1,37 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "svitch"
+version = "0.1.5"
+description = "Privacy-first AI infrastructure — PII detection, redaction, agent audit trails, and compliance reports for teams building AI on sensitive data."
+readme = "README.md"
+license = { text = "Apache-2.0" }
+requires-python = ">=3.10"
+keywords = ["pii", "privacy", "llm", "security", "compliance", "gdpr", "dpdp", "hipaa", "aadhaar", "pan", "agent", "audit"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: Apache Software License",
+    "Programming Language :: Python :: 3",
+    "Topic :: Security",
+    "Topic :: Software Development :: Libraries",
+]
+dependencies = []
+
+[project.optional-dependencies]
+openai = ["openai>=1.0.0"]
+anthropic = ["anthropic>=0.25.0"]
+langchain = ["langchain-core>=0.2.0"]
+otel = ["opentelemetry-api>=1.0.0"]
+all = ["openai>=1.0.0", "anthropic>=0.25.0", "langchain-core>=0.2.0", "opentelemetry-api>=1.0.0"]
+
+[project.urls]
+Homepage = "https://svitch.ai"
+Documentation = "https://svitch.ai/dpdp"
+Repository = "https://github.com/koushiknarendra/svitch"
+Issues = "https://github.com/koushiknarendra/svitch/issues"
+
+[tool.hatch.build.targets.wheel]
+packages = ["svitch", "svitch_tracer"]

svitch-0.1.5/svitch/__init__.py

@@ -0,0 +1,14 @@
+from .shield import detect, redact, SvitchResult, Entity
+from .wrap import wrap
+from .router import Router, RouteResult
+from .cache import stabilize, StabilizeResult
+
+__all__ = ["detect", "redact", "wrap", "Router", "RouteResult", "SvitchResult", "Entity", "stabilize", "StabilizeResult"]
+__version__ = "0.1.5"
+
+
+def __getattr__(name: str):
+    if name == "SvitchCallbackHandler":
+        from .langchain import SvitchCallbackHandler
+        return SvitchCallbackHandler
+    raise AttributeError(f"module 'svitch' has no attribute {name!r}")

svitch-0.1.5/svitch/cache.py

@@ -0,0 +1,130 @@
+"""
+Prefix Stabilizer — normalize system prompts for LLM provider KV cache hits.
+
+Unstable system prompts (timestamps, session IDs, request tokens) cause a
+cache miss on every single call, paying full input token cost each time.
+Stabilizing the prefix lets the provider serve from its KV cache:
+
+  Anthropic — cache-read tokens cost ~10% of normal input token price.
+  OpenAI    — prompts > 1024 tokens are auto-cached when the prefix is stable.
+
+Latency impact: 30-50% lower TTFT on repeated agent calls with the same
+system prompt structure.
+
+This is also a compliance signal: timestamps and session IDs embedded in
+system prompts are inadvertent per-request data leaks. `StabilizeResult.removed`
+surfaces exactly what was replaced so the audit trail knows what changed.
+"""
+from __future__ import annotations
+
+import hashlib
+import re
+from dataclasses import dataclass, field
+
+# (label, compiled pattern, stable placeholder)
+_UNSTABLE: list[tuple[str, re.Pattern, str]] = [
+    (
+        "timestamp_iso",
+        re.compile(
+            r"\d{4}-\d{2}-\d{2}[T ]\d{2}:\d{2}(:\d{2})?(\.\d+)?(Z|[+-]\d{2}:?\d{2})?",
+        ),
+        "[TIMESTAMP]",
+    ),
+    (
+        "timestamp_date_phrase",
+        re.compile(
+            r"(today is|current date[:\s]+|as of[:\s]+)\s*[\w,]+ \d{1,2}[,\s]+\d{4}",
+            re.IGNORECASE,
+        ),
+        "[DATE_PHRASE]",
+    ),
+    (
+        "unix_timestamp",
+        re.compile(r"\b1[6-9]\d{8}\b"),
+        "[UNIX_TS]",
+    ),
+    (
+        "uuid",
+        re.compile(
+            r"\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b",
+            re.IGNORECASE,
+        ),
+        "[UUID]",
+    ),
+    (
+        "session_id",
+        re.compile(r"\bsession[_-]?id[:\s]+[\w\-]{8,}\b", re.IGNORECASE),
+        "[SESSION_ID]",
+    ),
+    (
+        "request_id",
+        re.compile(r"\brequest[_-]?id[:\s]+[\w\-]{8,}\b", re.IGNORECASE),
+        "[REQUEST_ID]",
+    ),
+    (
+        "hex_id",
+        re.compile(r"\b[0-9a-f]{24,}\b"),
+        "[HEX_ID]",
+    ),
+]
+
+
+@dataclass
+class StabilizeResult:
+    text: str           # normalized system prompt — safe to send to the provider
+    original: str       # unmodified original — preserved for audit
+    hash: str           # SHA-256 prefix of `text` — use for cache-hit rate tracking
+    removed: list[str]  # pattern labels replaced (e.g. ["timestamp_iso", "uuid"])
+    changed: bool       # True when at least one substitution occurred
+
+
+def stabilize(system: str) -> StabilizeResult:
+    """
+    Normalize a system prompt to maximize LLM provider KV cache hits.
+
+    Replaces per-request variables (timestamps, UUIDs, session/request IDs)
+    with stable placeholders. The original is preserved in `StabilizeResult.original`
+    so callers can log it to the audit trail without losing information.
+
+    Args:
+        system: Raw system prompt string.
+
+    Returns:
+        StabilizeResult with normalized text, SHA-256 prefix hash, and a list
+        of pattern labels that were replaced.
+    """
+    if not system:
+        return StabilizeResult(text=system, original=system, hash="", removed=[], changed=False)
+
+    text = system
+    removed: list[str] = []
+
+    for label, pattern, placeholder in _UNSTABLE:
+        new_text, n = pattern.subn(placeholder, text)
+        if n:
+            text = new_text
+            removed.append(label)
+
+    digest = hashlib.sha256(text.encode()).hexdigest()[:16]
+    return StabilizeResult(
+        text=text,
+        original=system,
+        hash=digest,
+        removed=removed,
+        changed=bool(removed),
+    )
+
+
+def anthropic_cache_block(text: str) -> list[dict]:
+    """
+    Wrap stabilized text in an Anthropic prompt-caching block.
+
+    Pass the returned list as the `system` parameter to `client.messages.create()`.
+    Anthropic caches tokens at this breakpoint; subsequent requests with the
+    same prefix pay ~10% of normal input token cost (minimum 1024 tokens).
+
+    Example:
+        system_blocks = anthropic_cache_block(stabilize(raw_system).text)
+        client.messages.create(model=..., system=system_blocks, messages=...)
+    """
+    return [{"type": "text", "text": text, "cache_control": {"type": "ephemeral"}}]

svitch-0.1.5/svitch/classify.py

@@ -0,0 +1,65 @@
+"""
+Prompt complexity classifier — decides "fast", "default", or "complex".
+Used by the Router to pick the right model tier automatically.
+"""
+from __future__ import annotations
+
+_COMPLEX_KEYWORDS = frozenset([
+    "analyze", "analyse", "explain", "compare", "evaluate", "assess",
+    "research", "investigate", "comprehensive", "detailed", "thorough",
+    "implement", "architecture", "strategy", "design",
+    "write a", "create a", "build a", "develop", "refactor", "debug",
+    "step by step", "in depth", "elaborate", "translate",
+    "legal", "medical", "financial", "compliance", "audit",
+    "summarize", "generate code", "write code", "essay",
+])
+
+_SIMPLE_PREFIXES = ("what is ", "who is ", "when did ", "where is ",
+                    "define ", "list ", "name ")
+
+
+def _token_estimate(text: str) -> int:
+    return max(1, int(len(text.split()) * 1.3))
+
+
+def classify(messages: list[dict]) -> str:
+    """
+    Returns "fast", "default", or "complex".
+
+    fast    — simple lookup, extraction, short Q&A       → cheapest model
+    default — standard reasoning, explanation             → standard model
+    complex — deep analysis, long-form, code generation  → strongest model
+    """
+    user_turns = [m for m in messages if m.get("role") == "user"]
+    user_text  = " ".join(m.get("content", "") for m in user_turns).lower().strip()
+    total_tokens = sum(_token_estimate(m.get("content", "")) for m in messages)
+
+    score = 0
+
+    # Token volume is the strongest signal
+    if total_tokens > 600:
+        score += 3
+    elif total_tokens > 200:
+        score += 1
+
+    # Multi-turn context
+    if len(user_turns) > 2:
+        score += 1
+
+    # Complex keyword hit
+    for kw in _COMPLEX_KEYWORDS:
+        if kw in user_text:
+            score += 2
+            break
+
+    # Simple prefix — negative signal
+    for prefix in _SIMPLE_PREFIXES:
+        if user_text.startswith(prefix):
+            score -= 1
+            break
+
+    if score <= 1:
+        return "fast"
+    if score <= 4:
+        return "default"
+    return "complex"