Career Chat¶

This is a simple implementation of RAG to chat over my career history.

Note:

  • VectorDB is not used. FAISS (vector index) used because career data changes less frequently and also keeps it lightweight since index files can be loaded directly from the file-system
  • Local LLM took too long to compute, hence this notebook has been revised to use OpenAI LLM

Setup¶

With the following contents of pyproject.toml created by uv:

[project]
name = "career-chat"
version = "0.1.0"
description = "Career Chat with RAG - chunks and FAISS indexes"
readme = "README.md"
requires-python = ">=3.12"
dependencies = [
    "faiss-cpu>=1.11.0.post1",
    "langchain-openai>=0.3.28",
    "langchain-text-splitters>=0.3.8",
    "markdown-it-py>=3.0.0",
    "numpy>=2.3.1",
    "pypandoc>=1.15",
    "python-dotenv>=1.1.1",
    "sentence-transformers>=5.0.0",
]

[dependency-groups]
dev = [
    "black>=25.1.0",
]

Run the following to resolve depedencies

$ uv sync --all-groups
Resolved 76 packages in 7ms
Audited 57 packages in 0.36ms
In [1]:
import os
import re
from langchain_text_splitters import MarkdownHeaderTextSplitter
import json
import numpy as np
from typing import List, Dict, Tuple
from sentence_transformers import SentenceTransformer
import faiss
import openai
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage

Birds Eye View¶

High level depiction of the entire pipelien

In [13]:
from IPython.display import Image, display
import base64
import requests

graph_definition = """
flowchart TD
 subgraph Preprocessing["Text Preprocessing"]
        A
        B
        BC
  end
 subgraph Indexing["Indexing Pipeline"]
        C
        D
  end
 subgraph Retrieval["Retrieval"]
        E
        EC
  end
 subgraph Generation["Generation (via LLM)"]
        F
        G
  end
    A["Notion Markdown"]:::bold -- "manual-export" --> B["Semantic chunking"]
    B -- chunks --> BC
    BC[Chunks] --> C
    C["<b>Embedding Model</b>"]:::bold --> D["<b>FAISS Vector Index</b>"]
    U["User Query"] --> F
    EC --> |"use chunks as context<br>in LLM prompt<br>(aka Augmentation)"|F["Prompt Construction"]
    style EC stroke-dasharray: 3 3,stroke:#666,stroke-width:1px,fill:#f5f5f5,radius:5px
    BC -.-> EC["Retrieved Chunks"]
    D --> E["Retrieve Top-K Documents"]
    E --> EC["Retrieved Chunks"]
    F --> G["LLM with Context"]
    G --> H["Final Response"]
     D:::component
     C:::component
     B:::component
     A:::component
     E:::component
     F:::component
     G:::component
     H:::component
     U:::component
    classDef component fill:#e6f7ff,stroke:#4dabf7
    style A fill:#bbb,stroke:#333
    style B fill:#bbf,stroke:#333
    style C fill:#f99,stroke:#333
    style D fill:#ff9,stroke:#333
    style E fill:#bbf,stroke:#333
    style F fill:#f9f,stroke:#333
    style G fill:#9f9,stroke:#333
    style H fill:#f9a,stroke:#333

"""

def render_mermaid(graph):
    graphbytes = graph.encode("utf-8")
    base64_bytes = base64.b64encode(graphbytes)
    base64_string = base64_bytes.decode("ascii")
    img_url = "https://mermaid.ink/img/" + base64_string
    display(Image(url=img_url))

render_mermaid(graph_definition)
No description has been provided for this image

Data Prep¶

My content are in Notion where I have structured it in a specific way. In this example, I have exported the markdown content for the Notion page and will process it to chunk them.

In [2]:
class MarkdownCVChunker:
    """
    A specialised chunker for CV in the following format:
    ```
        # Career Profile
        # Company
        <optional brief>
        ## Role <dates>
        <optional brief>
        ### Tech work
        <optional brief>
        Contains tabular information with columns:
        Job Title |	Work/Project |	Technology Used |	Skills Applied |	Description	| Highlights
        ### <Optional> Non Tech work
        <optional brief>
        Contains tabular information with columns:
        Job Title |	Work/Project |	Skills Applied |	Description	| Highlights
        ### <Optional> Other insights or remarks

        # <Optional> Personal Mini Project & Hobbies
        list of projects

        # Training/Certifications/Courses
        bullet list

        #Education
        Uni degrees etc
    ```
    """

    def __init__(self):
        self.splitter = MarkdownHeaderTextSplitter(
            headers_to_split_on=[
                ("#", "H1"),
                ("##", "H2"),
                ("###", "H3"),
            ],
            strip_headers=True,
        )  # Set strip_headers to True

        self.table_regex = re.compile(
            r"^\s*\|.*\|\s*\n" r"^\s*\|[-|: ]+\|?\s*\n" r"(?:^\s*\|.*\|\s*\n?)*",
            re.MULTILINE,
        )

    def _parse_table(self, table_markdown: str) -> list[dict]:
        """Parses a Markdown table into a list of dictionaries."""
        lines = table_markdown.strip().split("\n")
        if len(lines) < 2:
            return []
        headers = [h.strip() for h in lines[0].strip().strip("|").split("|")]
        rows = []
        for line in lines[2:]:
            if not line.strip().startswith("|"):
                continue
            cells = [c.strip() for c in line.strip().strip("|").split("|")]
            if len(cells) == len(headers):
                rows.append(dict(zip(headers, cells)))
        return rows

    def process_cv(self, markdown_text: str) -> list[str]:
        """
        Processes the full CV into a list of clean, non-duplicated
        Markdown chunks.
        """
        final_chunks = []
        # The splitter gives us docs where page_content is the text
        # between headers, and the headers are in metadata.
        initial_docs = self.splitter.split_text(markdown_text)

        for doc in initial_docs:
            content = doc.page_content.strip()

            # Check if this specific piece of content contains a table
            table_match = self.table_regex.search(content)

            if table_match:
                # This is a "Table Container" chunk 
                # Its only job is to provide rows for granular chunks.

                # Extract context from metadata
                h1 = doc.metadata.get("H1", "").strip()
                h2 = doc.metadata.get("H2", "").strip()
                h3 = doc.metadata.get("H3", "").strip()

                table_str = table_match.group(0)
                table_rows = self._parse_table(table_str)

                for row in table_rows:
                    context_header = f"Regarding my work at **{h1}** as **{h2}**"
                    if h3:
                        context_header += f" ({h3}):\n\n"
                    else:
                        context_header += ":\n\n"

                    chunk_md = context_header
                    for key, value in row.items():
                        if value and key:
                            chunk_md += f"- **{key.strip()}:** {value.strip()}\n"
                    final_chunks.append(chunk_md)

                # Check for any text that was *not* the table
                # in this container, like an introductory sentence.
                non_table_text_in_chunk = self.table_regex.sub("", content).strip()
                if non_table_text_in_chunk:
                    # Prepend the context headers to this narrative text as well
                    context_header = ""
                    if h1:
                        context_header += f"# {h1}\n"
                    if h2:
                        context_header += f"## {h2}\n"
                    if h3:
                        context_header += f"### {h3}\n"
                    final_chunks.append(context_header + non_table_text_in_chunk)

            elif content:
                # This is a "Narrative Chunk"
                # It has no table. It's a good semantic unit on its own.
                # Reconstruct its full context from metadata for clarity.
                context_header = ""
                if "H1" in doc.metadata:
                    context_header += f"# {doc.metadata['H1']}\n"
                if "H2" in doc.metadata:
                    context_header += f"## {doc.metadata['H2']}\n"
                if "H3" in doc.metadata:
                    context_header += f"### {doc.metadata['H3']}\n"

                final_chunks.append(context_header + content)

        return final_chunks

Test it¶

The markdown file has been placed in cv-data

In [3]:
!tree cv-data/

cv-data/
└── cv-2025-06-01.md

1 directory, 1 file
In [4]:
# Initialize the chunker
chunker = MarkdownCVChunker()

# Process the CV to get Markdown chunks
cv_source_markdown_file_name = "cv-2025-06-01.md"
with open(f"cv-data/{cv_source_markdown_file_name}", "r", encoding="utf-8") as f:
    markdown_cv_text = "\n".join(line for line in f)
markdown_chunks = chunker.process_cv(markdown_cv_text)

# --- Output and Verification ---
print(
    f"✅ Successfully split the CV into {len(markdown_chunks)} Markdown chunks.\n"
)
from pprint import pprint

for c in markdown_chunks:
    print("-" * 20)
    pprint(f"{c[:53]} ... {c[-50:]}")

✅ Successfully split the CV into 56 Markdown chunks.

--------------------
('# Career Profile\n'
 'Overall, I learn quick, take on chal ...  lead by example, get shit done and '
 'provide value.')
--------------------
('# Healthdirect Australia\n'
 '📍Sydney, Australia  \n'
 'Sr. Sof ... of Engineering  \n'
 'Dates: December 4, 2017 - Present')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... compliant to '
 'organisational AWS design guidelines\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... e fees. '
 '**Trailblazer in LLM based AI adoption**.\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ard and the '
 'service funders from the Commonwealth\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... imilar '
 'low-friction transition to primary system.\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... s: AWS Platform, '
 'NHSD, Contact Centre Development\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ments were now '
 'aligned, enhancing standardisation\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ...  the highest '
 'across all teams - tech and non-tech\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... es\n'
 '- **Highlights:** Can’t divulge. Confidential.\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ctive. Candidate '
 'selection as positively affected\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ered in relatively '
 'short period and within budget\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ights:** Enabled '
 'team to deliver at a rapid pace.\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... MS Teams with '
 'custom features not available COTS)\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... mal **saving 50K '
 'USD per annum in license fees.**\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... abilities that '
 'often required custom development.\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... : July 6, 2020 → '
 'February 5, 2021** (Tech work):\n'
 '\n')
--------------------
('# Healthdirect Australia\n'
 '## Principal Engineer  - Dat ... ect saved annual recurring license fees by '
 '50K USD')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ption:** Developed '
 'various backend API components\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... d Typescript into '
 'the team development workflow**\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... :** '
 '**Trailblazer** in successful NLU application\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ectives. Was '
 'rolled out for various NSW hospitals\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... ough vendors like '
 'HealthEngine, HotDoc and others\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... to pave its way '
 'into adoption in the organisation\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... all solutions '
 'applying best-in-class principles**\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... rential treatment '
 'towards granting new projects**\n')
--------------------
('Regarding my work at **Healthdirect Australia** as ** ... : December 4, 2017 '
 '→ July 3, 2020** (Tech work):\n'
 '\n')
--------------------
('# Healthdirect Australia\n'
 '## Sr. Software Developer  - ... sign with advanced cloudformation based '
 'automation')
--------------------
('# Insight Timer\n'
 '📍Sydney, Australia  \n'
 'Sr. Software Eng ... ineer  \n'
 'Dates: February 4, 2017 → December 4, 2017')
--------------------
('Regarding my work at **Insight Timer** as **** (Tech  ... ucing cost by '
 'lowering dependency on AWS services\n')
--------------------
('Regarding my work at **Insight Timer** as **** (Tech  ...  adopted Kotlin '
 'over java for new implementations\n')
--------------------
('# Insight Timer\n'
 '### Other highlights or remarks\n'
 '- Go  ... ng, REST API design, RxJava/Java and System design')
--------------------
('# Autodesk Canada\n'
 '📍Toronto, Canada  \n'
 'Sr. Java Develop ... eloper  \n'
 'Dates: January 2, 2012 → January 18, 2017')
--------------------
('Regarding my work at **Autodesk Canada** as **** (Tec ... end supporting the '
 'popular Fusion360 product line\n')
--------------------
('Regarding my work at **Autodesk Canada** as **** (Tec ...  outputs across '
 'geo-distributed development teams\n')
--------------------
('Regarding my work at **Autodesk Canada** as **** (Non ... pplied:** '
 'Organisation, team working, recruitment\n')
--------------------
('Regarding my work at **Autodesk Canada** as **** (Non ... velocity '
 'frictions\n'
 '- **Skills Applied:** Outreach\n')
--------------------
('# Autodesk Canada\n'
 '### Other highlights or remarks\n'
 '- G ... remarks\n'
 '- Go to person for HATEOAS REST API design')
--------------------
('# University of Iowa\n'
 '📍Iowa City, Iowa, USA  \n'
 'Applicat ... oper  \n'
 'Dates: December 1, 2008 → December 15, 2011')
--------------------
('Regarding my work at **University of Iowa** as **** ( ... col, and CORS for '
 'creating well behaving web apps\n')
--------------------
('# University of Iowa\n'
 '### Tech work\n'
 'Exposed a web-appl ... t graduates in the development of jQuery mobile. |')
--------------------
('Regarding my work at **University of Iowa** as **** ( ... edge sharing\n'
 '- **Highlights:** Recruited speakers\n')
--------------------
('# University of Iowa\n'
 '### Other highlights or remarks\n'
 ' ... rapport with stakeholders during Business Analysis')
--------------------
('# ACT Incorporated\n'
 '📍Iowa City, Iowa, USA  \n'
 'Staff Anal ... alyst  \n'
 'Dates: October 1, 2007 → November 26, 2008')
--------------------
('Regarding my work at **ACT Incorporated** as **** (Te ... equirements '
 'gathering from resident Test Authors.\n')
--------------------
('Regarding my work at **ACT Incorporated** as **** (No ... ent CMS\n'
 '- **Description:** Requirements gathering\n')
--------------------
('# D2 Hawkeye Pvt Ltd\n'
 '📍Kathmandu, Nepal  \n'
 'Java Develop ... \n'
 'Java Developer  \n'
 'Dates: May 1, 2005 → May 1, 2006')
--------------------
('Regarding my work at **D2 Hawkeye Pvt Ltd** as **** ( ... ort engine became '
 'one of their flagship offerings\n')
--------------------
('Regarding my work at **D2 Hawkeye Pvt Ltd** as **** ( ... Created '
 'internal-use SOAP API endpoints for data.\n')
--------------------
('Regarding my work at **D2 Hawkeye Pvt Ltd** as **** ( ... with US staff\n'
 '- **Highlights:** Met key deadlines\n')
--------------------
('# Self Employed\n'
 '📍Kathmandu, Nepal  \n'
 'Visual Basic 6 De ...  Developer  \n'
 'Dates: December 1, 2004 → May 1, 2005')
--------------------
('Regarding my work at **Self Employed** as **** (Tech  ... rything in '
 'between, including support/maintenance\n')
--------------------
('Regarding my work at **Self Employed** as **** (Non T ... gathering, Price '
 'negotiation, Contract securement\n')
--------------------
('# Personal Mini Project & Hobbies\n'
 '- Q&A (RAG) text bo ... lajs-mastermind-game)  \n'
 '- Tech: ScalaJs, HTML, CSS')
--------------------
('# Training/Certifications/Courses\n'
 '- Generative AI wit ... unctional Programming in Haskell - January 4, 2016')
--------------------
('# Education\n'
 '- Masters - Computer Science  \n'
 'Maharishi  ... ieeexplore.ieee.orgstampstamp.jsparnumber1310247))')

Save chunks as json¶

Now let's save these chunks as json

In [5]:
data_to_save = [
    {"text": chunk} for chunk in markdown_chunks
]  # Placeholder for embedding step
output_filename = os.path.join(
    "prepared_data_faiss", f"{cv_source_markdown_file_name}.chunks.json"
)
# Create directory structure if it doesn't exist
os.makedirs("prepared_data_faiss", exist_ok=True)
with open(output_filename, "w") as f:
    json.dump(data_to_save, f, indent=2)
print(
    f"💾 Mock data saved to '{output_filename}'. Next step would be to add embeddings."
)

💾 Mock data saved to 'prepared_data_faiss/cv-2025-06-01.md.chunks.json'. Next step would be to add embeddings.

The career data is naturally grouped into semantic elements. Hence adopted this chunking strategy instead of a RecursiveCharacterTextSplitter.

Embedding¶

Next we will add these chunks into the vector index so that they can be semnatically retrieved later

In [6]:
class CVFaissEmbedder:
    def __init__(self, model_name: str = "sentence-transformers/all-MiniLM-L6-v2"):
        self.model = SentenceTransformer(model_name)
        self.index = None
        self.chunks = []
        self.dimension = 384  # dimension for "sentence-transformers/all-MiniLM-L6-v2"

    def load_chunks(self, chunks_json_file: str) -> List[Dict]:
        with open(chunks_json_file, "r") as f:
            self.chunks = json.load(f)
        return self.chunks

    def create_embeddings(self) -> np.ndarray:
        """Create embeddings for all chunks"""
        embeddings = self.model.encode(self.chunks, convert_to_numpy=True)
        return embeddings

    def build_index(self, embeddings: np.ndarray):
        """Build FAISS index"""
        self.index = faiss.IndexFlatIP(self.dimension)
        faiss.normalize_L2(embeddings)
        self.index.add(embeddings)

    def save_index(self, file: str = "cv_index.faiss"):
        faiss.write_index(self.index, file)
        print(f"Saved index to {file}")

    def load_index(self, file: str = "cv_index.faiss"):
        self.index = faiss.read_index(file)
        print(f"Loaded index from {file}")

    def search(self, query: str, num_chunks_to_retrieve: int = 3) -> List[Dict]:
        query_embeds = self.model.encode([query], convert_to_numpy=True)
        faiss.normalize_L2(query_embeds)
        # search
        scores, indices = self.index.search(query_embeds, num_chunks_to_retrieve)
        # map to chunks from indices
        results = []
        for idx, score in zip(indices[0], scores[0]):
            chunk = {}
            if idx < len(self.chunks):  # safety check
                chunk["content"] = str(self.chunks[idx])  # .copy()
                chunk["relevance_score"] = float(score)
                results.append(chunk)
        return results

Perform embedding¶

Now let's do the embedding

In [7]:
cv_embeds = CVFaissEmbedder()
cv_embeds.load_chunks(output_filename)
embeddings = cv_embeds.create_embeddings()
print(embeddings)

print("Now building FAISS index")
cv_embeds.build_index(embeddings)
embeddings_filename=f"{output_filename}.faiss"
cv_embeds.save_index(embeddings_filename)

[[-0.01746857 -0.01386396  0.00048448 ... -0.03198443 -0.00998253
   0.00830058]
 [-0.05188219  0.02842885  0.04710363 ... -0.06577165  0.06617153
   0.05583187]
 [-0.09246843  0.09975132 -0.04714889 ... -0.06883582  0.01592954
   0.05648626]
 ...
 [-0.10956752 -0.02698764 -0.00150927 ... -0.02031867  0.0445997
   0.04409875]
 [-0.05473209 -0.06764245  0.05323834 ...  0.01342495 -0.00667846
  -0.00751785]
 [-0.05910152  0.06252933 -0.0561499  ... -0.01943959 -0.02822687
   0.0874305 ]]
Now building FAISS index
Saved index to prepared_data_faiss/cv-2025-06-01.md.chunks.json.faiss

Test semantic retrieval¶

Now let's test vector index retrieval to see how well it matches

In [8]:
print("Work history:\n")
pprint(cv_embeds.search('work history'))
print("-" * 20)
print("Background:\n")
pprint(cv_embeds.search('Background'))
print("-" * 20)

Work history:

[{'content': "{'text': 'Regarding my work at **Healthdirect Australia** as "
             '**Sr. Software Developer  - Dates: December 4, 2017 → July 3, '
             '2020** (Tech work):\\n\\n- **Job Title:** Sr. Software '
             'Engineer\\n- **Work/Project:** CMS API development\\n- '
             '**Technology Used:** Java 8, Alfresco, PoolParty (RDF)\\n- '
             '**Skills Applied:** Integration\\n- **Description:** Developed '
             "various backend API components\\n'}",
  'relevance_score': 0.41815584897994995},
 {'content': "{'text': 'Regarding my work at **Autodesk Canada** as **** (Non "
             'Tech work):\\n\\n- **Work/Project:** Cross-geo hackathons\\n- '
             '**Description:** Organised 2 such yearly hackathons\\n- **Skills '
             "Applied:** Organisation, team working, recruitment\\n'}",
  'relevance_score': 0.4164963662624359},
 {'content': "{'text': 'Regarding my work at **Healthdirect Australia** as "
             '**Sr. Software Developer  - Dates: December 4, 2017 → July 3, '
             "2020** (Tech work):\\n\\n'}",
  'relevance_score': 0.40159571170806885}]
--------------------
Background:

[{'content': "{'text': '# Autodesk Canada\\n### Other highlights or "
             "remarks\\n- Go to person for HATEOAS REST API design'}",
  'relevance_score': 0.21256308257579803},
 {'content': "{'text': '# Personal Mini Project & Hobbies\\n- Q&A (RAG) text "
             'bot that answers questions based on my CV (April 15, 2025)  \\n- '
             'Atlassian Forge Plugin (December 23, 2024):  \\n- A plugin for '
             'Confluence to create/update/edit Lean Canvas  \\n- Tech: React, '
             'Typescript, mobx, Atlassian Forge  \\n- Simple Agentic '
             'researcher (February 11, 2024 - Github link: '
             'https://github.com/ffos/llm-agentic-search)  \\n- Supervisor '
             'agent accepted user query and produced a research report '
             'delegating to sub-agents to using DuckDuckGo as “tool”  \\n- '
             'Tech: Langchain (Python), OpenAI (LLM)  \\n- Woodworking and '
             'indoor furniture making (Attained TAFE Cert II - November 1, '
             '2023 )  \\n- “Mastermind” game running on browser (November 27, '
             '2016 - Github link: '
             'https://github.com/ffos/scalajs-mastermind-game)  \\n- Tech: '
             "ScalaJs, HTML, CSS'}",
  'relevance_score': 0.2050330638885498},
 {'content': "{'text': '# Education\\n- Masters - Computer Science  "
             '\\nMaharishi International University  \\nIowa, USA  \\nFebruary '
             '1, 2006 → February 1, 2008  \\n- Bachelors - Computer '
             'Engineering  \\nInstitute of Engineering, Tribhuvan University  '
             '\\nLalitpur, Nepal  \\nFebruary 1, 2000 → February 1, 2005  '
             '\\nFinal project was IEEE CSIDC 2004 - World 10th '
             '([CSIDC-Link-1](https://archive.md/OWLlZ), '
             "[CSIDC-Link-2](https://archive.org/details/httpsieeexplore.ieee.orgstampstamp.jsparnumber1310247))'}",
  'relevance_score': 0.1774899661540985}]
--------------------

Use LLM to return results¶

  1. Find matches for query using FAISS (the query needs to be embedded too so they lookup in the same vector space)
  2. Use the match to build the context for LLM to respond with
In [9]:
from langchain_core.messages import HumanMessage, SystemMessage
load_dotenv()
print(f"Open AI API Key string of length {len(os.getenv('OPENAI_API_KEY'))} now loaded")
llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0.7, max_tokens=500)

def chat(query:str) -> dict[str,any]:
    context_chunks = cv_embeds.search(query, num_chunks_to_retrieve=20)
    context = "\n\n".join(c["content"] for c in context_chunks)
    instruction = (
        "You are an AI assistant representing a professional's CV."
        "Answer questions based ONLY on the provided CV information within <CONTEXT> tags."
        "Be concise, accurate and professional."
        "If the information is not in the context, say so politely."
        f"<Context>\n{context}\n</Context>\n\n"
        "Response:"
    )
    messages = [
        SystemMessage(content=instruction),
        HumanMessage(content=query)
    ]
    response = llm.invoke(messages)
    return response.content

Open AI API Key string of length 164 now loaded
In [10]:
pprint(chat("Tell me about yourself"))

('Overall, I am quick to learn, embrace challenges, and lead by example. I '
 'adapt quickly, provide value, and am optimistic and friendly in my approach.')
In [11]:
output = chat("list all the company names you worked along with date ranges. Output json.")
parsed_json = json.loads(output)
print(json.dumps(parsed_json, indent=2))

{
  "companies": [
    {
      "name": "Healthdirect Australia",
      "date_range": "December 4, 2017 - Present"
    },
    {
      "name": "Self Employed",
      "date_range": "December 1, 2004 - May 1, 2005"
    },
    {
      "name": "University of Iowa",
      "date_range": "December 1, 2008 - December 15, 2011"
    },
    {
      "name": "Autodesk Canada",
      "date_range": "January 2, 2012 - January 18, 2017"
    },
    {
      "name": "ACT Incorporated",
      "date_range": "October 1, 2007 - November 26, 2008"
    },
    {
      "name": "D2 Hawkeye Pvt Ltd",
      "date_range": "May 1, 2005 - May 1, 2006"
    },
    {
      "name": "Insight Timer",
      "date_range": "February 4, 2017 - December 4, 2017"
    }
  ]
}
In [12]:
output = chat("List all the company names you worked and the technologies used in each. Output json.")
parsed_json = json.loads(output)
print(json.dumps(parsed_json, indent=2))

{
  "work_experience": [
    {
      "company_name": "Healthdirect Australia",
      "technologies_used": [
        "Java 8",
        "Alfresco",
        "PoolParty (RDF)",
        "NodeJs/Typescript",
        "MithrilJs",
        "Snips-AI (Python)",
        "AWS Lambda",
        "API Gateways",
        "Cloudformation"
      ]
    },
    {
      "company_name": "D2 Hawkeye Pvt Ltd",
      "technologies_used": [
        "Java & Servlets",
        "Axis2",
        "Apache POI",
        "Oracle 10g",
        "SOAP"
      ]
    },
    {
      "company_name": "University of Iowa",
      "technologies_used": [
        "Java",
        "Jersey Framework",
        "Stripes Framework",
        "Hibernate ORM",
        "Oracle DB",
        "JSP",
        "jQuery",
        "prototype-js",
        "scriptaculous-js",
        "Groovy",
        "Oracle DB",
        "jQuery mobile"
      ]
    },
    {
      "company_name": "Autodesk Canada",
      "technologies_used": [
        "Java 8",
        "Postgres",
        "Jersey",
        "Antlr4",
        "Spring",
        "Hibernate",
        "ElasticSearch",
        "JBehave",
        "Angular JS",
        "Apache Tomcat",
        "Linux deployments on AWS EC2s"
      ]
    }
  ]
}