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.

Every agent in this workflow is a plain Python function that constructs an f-string prompt, invokes the LLM, and returns a dict with one output field. Because all configuration lives directly in source code rather than a config file, customizing the workflow means editing Python — no framework-specific abstractions to learn. This page walks you through the three most common changes: editing a prompt, adding a new state field, and adjusting LLM behavior.

​
Changing an agent’s prompt

Each agent builds its prompt as an f-string and passes it to llm.invoke(). To change what an agent does, edit the text of that f-string. Here is the full prompt from resume_parser_agent.py: 
from llm import llm

def resume_parser_agent(state):

    resume_text = state["resume_text"]

    response = llm.invoke(
        f"""
        You are a Resume Parsing Agent.

        Extract structured candidate information from this resume.

        Resume:
        {resume_text}

        Extract:
        - candidate_name
        - email
        - phone
        - skills
        - frameworks
        - experience
        - projects
        - education
        - certifications
        - github
        - linkedin
        - current_role
        - seniority_level

        Return structured JSON only.
        """
    )

    return {
        "candidate_profile": response.content
    }
To add a new extraction field — for example, languages_spoken — append it to the Extract: list: 
        Extract:
        - candidate_name
        - email
        ...
        - seniority_level
        - languages_spoken   # <-- add here
To change the output format from JSON to a numbered list, replace Return structured JSON only. with your preferred instruction. Keep in mind that downstream agents (such as matching and hr_interview) read candidate_profile and expect JSON, so changing the format requires updating those agents too.

​
Adding output fields

Follow these four steps any time you want a new agent to produce a new piece of data.
1

Add the field to HiringState

Open graph/state.py and add your new field to the HiringState TypedDict. All fields are Optional by default because total=False is set on the class.
# graph/state.py
class HiringState(TypedDict, total=False):
    # ... existing fields ...
    salary_benchmark: str   # <-- new field
2

Return the field from your agent

In your agent function, add the new key to the returned dict.
return {
    "salary_benchmark": response.content
}
LangGraph merges this dict into the shared state, so all subsequent agents will see salary_benchmark available.
3

Register the field with the supervisor

Open agents/supervisor.py and find the agents_info dict. Add your agent’s entry (or update an existing one) so the supervisor knows what the agent provides and requires. This is how the supervisor decides whether the required inputs are available before routing.
agents_info = {
    # ... existing entries ...
    "salary_research": {
        "purpose": "Researches market salary ranges for the role.",
        "requires": ["jd_analysis"],
        "provides": ["salary_benchmark"]
    },
}
4

Access the field in downstream agents or main.py

Any agent that runs after your new agent can read the field from state:
salary_benchmark = state.get("salary_benchmark", "")
To print it in main.py, add a branch to the output loop:
elif node == "salary_research":
    print(state.get("salary_benchmark"))

​
Adjusting LLM temperature

The temperature is set in llm.py: 
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gemma-3-4b-it",
    base_url="http://localhost:1234/v1",
    api_key="lm-studio",
    temperature=0.3,
    streaming=True
)
temperature controls randomness. The default 0.3 is a good balance for this workflow:
  • Lower temperature (0.0–0.2) — more deterministic output. Use this when agents must return valid JSON (such as resume_parser or evaluation). Malformed JSON causes the supervisor’s json.loads() call to raise an exception and route to finished early.
  • Higher temperature (0.5–0.9) — more creative, varied output. Use this if you want the interview agents to generate more diverse questions across runs. Be aware that higher temperature also increases the chance of malformed JSON.
Because all agents share the same llm instance, changing temperature in llm.py affects every agent. If you need different temperatures per agent, instantiate separate ChatOpenAI objects in individual agent files.

​
Changing the number of questions

The hr_interview_agent prompt currently asks for 15 HR questions. To change that, open agents/hr_interview_agent.py and edit this line: 
        Generate:
        - 15 HR questions    # <-- change the number here
        - evaluation criteria
        - red flags
The same pattern applies to technical_interview_agent.py and ceo_interview_agent.py — find the Generate: section and update the count.
Changing agent output field names requires updating every downstream agent that reads that field, plus the supervisor’s agents_info registry. For example, renaming candidate_profile to parsed_resume requires edits in matching_agent.py, hr_interview_agent.py, technical_interview_agent.py, ceo_interview_agent.py, candidate_research_agent.py, email_agent.py, and the supervisor’s requires lists.

Build docs developers (and LLMs) love

Get started for freeTalk to us