Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/vrashmanyu605-eng/Langchain_Interview_Multi_Agents_Flow/llms.txt

Use this file to discover all available pages before exploring further.

When you run the workflow, LangGraph passes a shared state dictionary through a series of specialist agents orchestrated by the supervisor. Each agent reads the fields it needs, performs its task using the local LLM, and writes its output back to state. The supervisor then inspects the updated state and decides which agent to invoke next. This continues until the email agent completes its draft and the supervisor routes to finished. You see each agent’s output printed to the terminal in real time as nodes complete — one node at a time, in streaming fashion.

​
Prepare and run the workflow

1

Prepare inputs

You need two files before running: a resume in PDF format and a job description as a plain .txt file. Update the path variables at the top of main.py to point to your local files.
resume_pdf_path = "path/to/resume.pdf"
jd_text_path = "path/to/jd.txt"
The extract_pdf_text helper (from src/services/pdf_parsing) reads the PDF and returns plain text. The JD file is opened and read directly.
2

Build the initial state dict

The workflow starts from a fixed initial state. This dict seeds the HiringState TypedDict with the raw documents and workflow control fields. Copy this block from main.py exactly — the supervisor expects task, resume_text, jd_text, workflow_stage, completed, and workflow_history to be present at startup.
initial_state = {
    "task": "Process candidate hiring workflow",
    "resume_text": resume_text,
    "jd_text": jd_text,
    "workflow_stage": "start",
    "completed": False,
    "workflow_history": []
}
3

Call app.stream() and iterate events

Pass initial_state to app.stream(). LangGraph yields one event dict per node execution. Iterate the outer loop to receive events and the inner loop to unpack the node name and its output state.
final_state = None

for event in app.stream(initial_state):
    for node, state in event.items():
        print(f"\n====================================")
        print(f"NODE: {node.upper()}")
        # ... print node-specific fields
        final_state = state
Each iteration of the outer for loop represents one agent completing its work. The inner event.items() loop unpacks the single key–value pair inside that event.
4

Read node outputs

Inside the inner loop, branch on the node name to print the relevant state field. Each agent writes exactly one output field; the supervisor writes next_agent. The complete output block from main.py is:
if node == "supervisor":
    print(f"Next Agent: {state.get('next_agent')}")
elif node == "resume_parser":
    print(state.get("candidate_profile"))
elif node == "jd_analysis":
    print(state.get("jd_analysis"))
elif node == "matching":
    print(state.get("matching_analysis"))
elif node == "candidate_research":
    print(state.get("research_analysis"))
elif node == "hr_interview":
    print(state.get("hr_questions"))
elif node == "technical_interview":
    print(state.get("technical_questions"))
elif node == "ceo_interview":
    print(state.get("ceo_questions"))
elif node == "evaluation":
    print(state.get("evaluation"))
elif node == "email":
    print(state.get("email_content"))

​
Complete main.py source

import fitz
from graph.workflow import app
from src.services.pdf_parsing import extract_pdf_text


resume_pdf_path = "path/to/your/resume.pdf"
jd_text_path = "path/to/your/jd.txt"

# Read Resume
resume_text = extract_pdf_text(resume_pdf_path)

# Read JD Text
with open(jd_text_path, "r", encoding="utf-8") as file:
    jd_text = file.read()


initial_state = {
    "task": "Process candidate hiring workflow",
    "resume_text": resume_text,
    "jd_text": jd_text,
    "workflow_stage": "start",
    "completed": False,
    "workflow_history": []
}


# RUN GRAPH

final_state = None


for event in app.stream(initial_state):

    for node, state in event.items():

        print(f"\n====================================")
        print(f"NODE: {node.upper()}")


        if node == "supervisor":
            print(f"Next Agent: {state.get('next_agent')}")

        elif node == "resume_parser":
            print(state.get("candidate_profile"))

        elif node == "jd_analysis":
            print(state.get("jd_analysis"))

        elif node == "matching":
            print(state.get("matching_analysis"))

        elif node == "candidate_research":
            print(state.get("research_analysis"))

        elif node == "hr_interview":
            print(state.get("hr_questions"))

        elif node == "technical_interview":
            print(state.get("technical_questions"))

        elif node == "ceo_interview":
            print(state.get("ceo_questions"))

        elif node == "evaluation":
            print(state.get("evaluation"))

        elif node == "email":
            print(state.get("email_content"))

        final_state = state

​
Streaming vs. batch

The workflow streams node-by-node: app.stream() yields control back to your loop after each agent finishes, so you see output progressively rather than waiting for the entire pipeline to complete. This is useful for long-running workflows because you can observe which agents succeed, which fields are populated, and whether the supervisor is routing correctly — all in real time. If you prefer to wait for the full result, you can call app.invoke(initial_state) instead, which returns the final state dict directly without intermediate events.

​
Node output reference

NodeState field printed
supervisornext_agent — the name of the next agent to invoke
resume_parsercandidate_profile — structured JSON of parsed resume fields
jd_analysisjd_analysis — structured analysis of the job description
matchingmatching_analysis — candidate-to-JD fit scoring and gaps
candidate_researchresearch_analysis — background research on the candidate
hr_interviewhr_questions — 15 HR questions with evaluation criteria
technical_interviewtechnical_questions — role-specific technical questions
ceo_interviewceo_questions — culture-fit and leadership questions
evaluationevaluation — final hire/no-hire recommendation
emailemail_content — drafted email to the candidate or hiring manager
Save final_state after the loop — it contains all agent outputs in one dict. You can serialize it to JSON for logging, pass it to another system, or inspect individual fields like final_state["evaluation"] without re-running the workflow.

Build docs developers (and LLMs) love

Get started for freeTalk to us