Skip to content

langchain.md

Latest commit

 

History

History
60 lines (46 loc) · 2.29 KB

langchain.md

File metadata and controls

60 lines (46 loc) · 2.29 KB

LangChain — .klickd v4.0.0 Integration

Framework: LangChain · pip install langchain langchain-openai SDK: pip install klickd==4.0.0

LangChain's prompt + chain primitives map cleanly to .klickd: the payload becomes the system message, the chain stays provider-agnostic.

Reference adapter

The reusable helpers live at examples/v4/integrations/langchain/klickd_langchain.py. They have no LangChain dependency at import time — langchain_* imports are deferred to build_chain(), so the prompt builder is safe to use in any environment.

Minimal example

from klickd import load_klickd
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

# 1. Decode the .klickd file
with open("examples/v4/personas/03-fullstack-developer-en.klickd", "rb") as f:
    payload = load_klickd(f.read())

# 2. Build the system prompt (use the helper for full SPEC §29 compliance)
from examples.v4.integrations.langchain.klickd_langchain import klickd_to_system_prompt
system_prompt = klickd_to_system_prompt(payload)

# 3. Standard LangChain chain
prompt = ChatPromptTemplate.from_messages([
    ("system", system_prompt),
    ("human", "{input}"),
])
chain = prompt | ChatOpenAI(model="gpt-4o")

print(chain.invoke({"input": "Let's pick up where we left off."}).content)

Notes

  • Always inject .klickd content in the system role unless the payload explicitly sets injection_target to user_message / both — in which case klickd_to_system_prompt() prepends the JSON Injection Guard automatically (SPEC §25.3).
  • Strip _-prefixed debug / benchmark keys before injection (SPEC §29) — the helper does this for you.
  • verification_gates are surfaced to the model as a single short instruction. Enforce the actual gate semantics in the host application, not in the LLM — the LLM is the agent, not the referee.
  • The helper supports both plain (encrypted: false) and encrypted envelopes: pass passphrase=... to load_klickd_path().

Provider-agnostic by design

Swap ChatOpenAI for ChatAnthropic, ChatGroq, or any LangChain-supported chat model. The .klickd payload is the portable state layer; the chain stays untouched.