A support ticket triage pipeline built with LangChain Expression Language (LCEL). Given a customer message, it returns a category, sentiment, keywords, an escalation decision and a draft reply.
triage/chains.pyhas the full pipeline.examples/has one short, runnable script per LCEL pattern, all using the same tickets and prompts.
flowchart LR
T(["Raw ticket"]) --> P["prepare_ticket<br/><i>RunnableLambda</i><br/>clean text, mask emails"]
P --> A
subgraph A ["analysis · RunnableParallel"]
direction TB
C["category<br/>prompt | model | parser"]
S["sentiment<br/>prompt | model | parser"]
K["keywords<br/>prompt | model | list parser"]
I["ticket<br/>itemgetter"]
end
A --> R{"route_ticket<br/><i>@chain</i>"}
R -- "negative and<br/>billing or technical" --> E(["Escalate to a human"])
R -- "otherwise" --> D["reply chain<br/>prompt | model | parser"]
D --> O(["Draft reply"])
Every box is a Runnable, so the whole pipeline supports invoke, batch, stream and get_graph() like any single step.
The five sample tickets in triage/data.py, run through the pipeline with the offline model:
| Ticket | Category | Sentiment | Keywords | Action |
|---|---|---|---|---|
| I was charged twice for my subscription on 2026-11-03… | billing | negative | charged, twice, subscription | 🚨 escalate |
| The mobile app keeps crashing when I open the reports tab. | technical | neutral | mobile, crashing, reports | ✉️ auto reply |
| Thanks for the great service! How do I change the email… | account | po 90F2 sitive | thanks, great, service | ✉️ auto reply |
| Do you offer a student discount? | other | neutral | student, discount | ✉️ auto reply |
| The export feature is broken again and I am really frustrated… | technical | negative | export, feature, broken | 🚨 escalate |
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"python examples/chain_decorator.pypytestBy default, everything runs on FakeTriageChatModel, so no key or network is needed. It's a real BaseChatModel, so invoke, batch, streaming, pipes and graphs are all LangChain's own code. Only the answers are made up, using simple keyword rules.
To use OpenAI, copy .env.example to .env, then set:
LLM_PROVIDER=openai
OPENAI_API_KEY=...
OPENAI_MODEL=gpt-4o-mini
All model creation goes through get_model() in triage/model.py, so another provider only needs a new branch there. The OpenAI path hasn't been run against the live API, so OPENAI_MODEL may need updating.
What each step of prompt | model | parser receives and returns. Rounded boxes are data, square boxes are Runnables.
flowchart LR
I(["dict<br/>{'ticket': '...'}"]) --> P["ChatPromptTemplate"]
P --> V(["ChatPromptValue<br/>list of messages"])
V --> M["Chat model"]
M --> AI(["AIMessage"])
AI --> O["StrOutputParser"]
O --> S(["str<br/>'billing'"])
Listed from simplest to most composed.
| Script | Shows |
|---|---|
model_inputs.py |
Model parameters, messages, chat and few-shot prompts |
output_parsers.py |
String, list and custom parsers |
basic_chain.py |
prompt | model | parser |
batch.py |
batch and max_concurrency |
streaming.py |
stream |
runnable_sequence.py |
Runnable and RunnableSequence |
passthrough_assign.py |
RunnablePassthrough and .assign |
graph.py |
get_graph(), ASCII and Mermaid |
parallel.py |
RunnableParallel |
parallel_into_chain.py |
Piping a parallel step onward, dict shorthand |
runnable_lambda.py |
RunnableLambda |
chain_decorator.py |
@chain and the full pipeline |
| I want to | Use |
|---|---|
| Connect steps | prompt | model | StrOutputParser() |
| Run on one input | chain.invoke(input) |
| Run on many inputs | chain.batch(inputs, config={"max_concurrency": 5}) |
| Get output as it's generated | for chunk in chain.stream(input): ... |
| Inspect a chain | chain.steps, chain.get_graph().print_ascii() |
| Add a key and keep the rest | RunnablePassthrough.assign(new_key=other_chain) |
| Run branches at the same time | RunnableParallel(a=chain_a, b=chain_b) or {"a": chain_a, "b": chain_b} in a pipe |
| Pick one key from a dict | itemgetter("key") |
| Use a function as a step | RunnableLambda(fn), or just step | fn |
| Use a function with branching logic | @chain |
Streaming stops being incremental at a RunnableLambda, which waits for its whole input unless it's written as a generator.
Written against LangChain 1.7. Older examples online may differ:
- Imports come from
langchain_coreandlangchain_openai. - Use
prompt | model | parserinstead ofLLMChain. DatetimeOutputParserisn't inlangchain-core, sotriage/parsers.pyhas a small custom parser instead.
The LangChain logo in .github/images is a trademark of LangChain, Inc. It comes from Wikimedia Commons, where it is published under the MIT license, and is used here only to show the project is built with LangChain.