Sanskrit-Mesh

Why are your AI agents wasting millions of tokens being polite to each other?

Sanskrit-Mesh is an AI-native bytecode compiler for multi-agent LLM pipelines. It intercepts the structured payloads that frameworks like AutoGen, LangChain, and CrewAI generate automatically — agent messages, memory objects, tool calls, system prompts — and compresses them into an ultra-dense Intermediate Representation (IR) inspired by Panini's Sanskrit grammar. On the other side, it decompiles back to perfect English with zero data loss.

Result: 55–77% token reduction on agent-generated payloads. Zero logic changes to your pipeline.

What V1 does and doesn't do

✅ Compresses structured agent payloads (JSON keys, framework boilerplate, error messages, system prompts, status fields, tool call structures)

✅ Works with AutoGen, LangChain, or any OpenAI-format API call

❌ Does not compress freeform human text — a user typing "deploy my app" saves nothing. The dictionary covers agent vocabulary, not natural conversation.

❌ Does not speed up inference or reduce model size

---

The Problem

Multi-agent systems (AutoGen, CrewAI, LangChain) auto-generate and transmit repetitive structured payloads on every step:

{
  "sender": "Agent A",
  "receiver": "Agent B",
  "intent": "Request Clarification",
  "context": {
    "status": "failed",
    "message": "I encountered the following error: NullPointerException: object reference not set. Please advise on how to proceed."
  }
}

232 characters. Sent hundreds of times per pipeline run. You never wrote this — your framework did.

The Solution

Sanskrit-Mesh compresses it to:

{"s":"|AgA|","r":"|AgB|","i":"|Prashna|","c":{"st":"|F|","m":"|E:| |ShunyaDosha|. |?|"}}

88 characters. Same meaning. 62% smaller.

Three compression layers running simultaneously:

1. Key minification — JSON keys shrunk to 1–3 chars ("sender" → "s", "intermediate_steps" → "is_")
2. Semantic IR dictionary — 200+ agent phrases mapped to dense Sanskrit tokens
3. Whitespace stripping — removes bloat agents auto-generate

---

Why Paninian Grammar?

Panini formalized Sanskrit grammar 2,500 years ago into the most concise, unambiguous linguistic rule system ever written. It encodes complex meaning in single dense constructs — exactly what machine-to-machine communication needs. MemoryError: out of memory becomes SmritiBhara. ConnectionError: failed to establish connection becomes BandhanDosha. Dense, unambiguous, lossless.

---

Installation

# Clone and use directly
git clone https://github.com/krishanumanna48-ctrl/sanskrit-mesh.git
cd sanskrit-mesh

No dependencies required for the base compiler and middleware. LangChain and AutoGen integrations require those packages installed separately.

PyPI release (pip install sanskrit-mesh) coming soon.

---

Quickstart

Universal — Works With Any OpenAI-Format API

from middleware import SanskritMeshMiddleware

middleware = SanskritMeshMiddleware()

# These messages contain agent-generated content — compresses well
messages = [
    {"role": "system",    "content": "You are a helpful, harmless, and honest assistant. Think step by step before answering. Always respond in JSON format."},
    {"role": "assistant", "content": "I will execute the tool to deploy. The deployment failed. Running again..."},
    {"role": "tool",      "content": "I encountered the following error: ConnectionError: failed to establish connection. Please advise on how to proceed."},
]

# NOTE: human freeform text (user messages) compresses minimally.
# The savings come from system prompts, assistant messages, and tool responses.
compressed = middleware.compress_messages(messages)

response = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=compressed
)

print(middleware.get_savings_report())

System Prompt Compression

System prompts are one of the best targets — they're repetitive, written by developers, and run on every single API call.

from middleware import SanskritMeshMiddleware

middleware = SanskritMeshMiddleware()

system_prompt = (
    "You are a helpful, harmless, and honest assistant. "
    "You are operating in a multi-agent environment. "
    "Think step by step before answering. "
    "Always respond in valid JSON. "
    "Your goal is to complete the assigned task efficiently."
)

compressed = middleware.compress_system_prompt(system_prompt)
# Result: |sys:hhh| |sys:multi| |sys:CoT| |sys:json+| |sys:goal|
# 315 chars → 74 chars. 76.5% smaller. Runs on every call.

LangChain Integration

from middleware import SanskritMeshLangChainCallback
from langchain_openai import ChatOpenAI

# One line — attaches to any LangChain LLM
callback = SanskritMeshLangChainCallback(verbose=True)
llm = ChatOpenAI(model="gpt-4o", callbacks=[callback])

# System prompts, memory, agent scratchpad all get compressed automatically
response = llm.invoke("Deploy the application.")

print(callback.get_session_report())

Compress LangChain memory directly:

compressed_memory = callback.compress_memory(memory.chat_memory.dict())

AutoGen Integration

from middleware import SanskritMeshAutoGenHook
import autogen

hook = SanskritMeshAutoGenHook(verbose=True)

planner  = autogen.ConversableAgent("PlannerAgent", ...)
executor = autogen.ConversableAgent("ExecutorAgent", ...)

# Register — all agent-to-agent messages compressed before transmission
planner.register_hook(
    hookable_method="process_message_before_send",
    hook=hook.compress_hook
)

# Compress full conversation history before passing to a new agent
compressed_history = hook.compress_conversation_history(
    planner.chat_messages[executor]
)

Raw Compiler

from compiler import SanskritMeshCompiler

compiler = SanskritMeshCompiler()

payload = {
    "intent": "Request Clarification",
    "message": "I encountered the following error: IndexError: list index out of range. Please advise on how to proceed."
}

compressed = compiler.compile_payload(payload)
# {'i': '|Prashna|', 'm': '|E:| |KramaBhanga|. |?|'}

restored = compiler.decompile_payload(compressed)
assert restored == payload  # 100% lossless — guaranteed

---

Live Benchmark

Run the benchmark suite on your machine:

python benchmark.py

Test your own payload:

python benchmark.py --payload '{"role": "system", "content": "You are a helpful assistant. Think step by step."}'

Test a JSON file:

python benchmark.py --file my_agent_payload.json

Real Benchmark Results (V1)

Actual numbers from python benchmark.py — verifiable on your machine.

Benchmark	Original	Compressed	Saving
Simple agent message	232 chars	88 chars	62.1%
System prompt	373 chars	87 chars	76.7%
LangChain memory / chat history	864 chars	379 chars	56.1%
Complex nested multi-agent payload	833 chars	374 chars	55.1%
ReAct agent scratchpad	656 chars	335 chars	48.9%
Worst case (zero dictionary matches)	245 chars	236 chars	3.7%

Real-world average on agent-generated traffic: ~59–62%

The worst case benchmark is deliberately adversarial — pure narrative text with zero agent vocabulary. That's not what Sanskrit-Mesh is built for. Real agent pipelines land between 55–77%.

---

What V1 Can Actually Save

Payload Type	Max Observed	Typical Range	Notes
System prompts	76.7%	60–77%	Best target — repetitive, developer-written
Simple agent messages	62.1%	55–65%	Framework-generated boilerplate
Multi-agent nested payloads	55–62%	50–65%	AutoGen / CrewAI message chains
ReAct scratchpads	48.9%	40–55%	Mixed agent + reasoning text
Human freeform text	~0–4%	0–8%	Not the target — key minification only

V1 ceiling: ~77%. Real-world average on agent pipelines: ~59%.

---

Cost Savings at Scale

Based on real benchmark averages (~59% compression, GPT-4o input pricing at $5/1M tokens):

Monthly API Calls	Avg Token Reduction	Monthly Savings
10,000	59%	~$7
100,000	59%	~$74
1,000,000	59%	~$740
10,000,000	59%	~$7,400

Assumes average 500 tokens/call on agent pipelines. Human chat apps will see lower savings.

---

For Local LLM Users (Low-End PCs)

If you run agent pipelines locally via Ollama or llama.cpp with models like Llama 3.2, Phi-3, or Mistral, Sanskrit-Mesh extends your effective context window on the structured parts of your conversation. A 4K context model running an AutoGen pipeline gets meaningfully more turns before hitting the limit.

This does not help if you're just doing casual chat — the savings only apply to agent-structured messages.

import ollama
from middleware import SanskritMeshMiddleware

middleware = SanskritMeshMiddleware()

# Works well when messages contain agent/tool structured content
messages = [...]
compressed = middleware.compress_messages(messages)

response = ollama.chat(model="llama3.2", messages=compressed)

---

Roadmap

• Core compiler with 200+ IR dictionary entries
• System prompt compression
• LangChain callback integration
• AutoGen hook integration
• Universal OpenAI-format middleware
• Live benchmark tool with cost reporting
• PyPI package release (pip install sanskrit-mesh)
• CrewAI integration
• Adaptive dictionary that learns from your own agent traffic
• Fine-tuned Sanskrit-Mesh-3B — a model that natively reads/writes IR (no decompilation needed)
• Human prompt compression via LLMLingua integration
• Ollama / llama.cpp plugin

---

License

MIT — free forever. Stop paying OpenAI to read your agents' polite greetings.