aniket-mish / llm-headline-generator

Simple retrieval system with hybrid search

Retrieval

System Architecture

Data Extraction

name	description	industry	category	url	headline	subhead	cta	founded	location	tags
GitLab	GitLab is the first single application for the entire DevOps lifecycle. Only GitLab enables Concurrent DevOps, unlocking organizations from the constraints of today’s toolchain. GitLab provides unmatched visibility, radical new levels of efficiency and comprehensive governance to significantly compress the time between planning a change and monitoring its effect. This makes the software lifecycle 200% faster, radically improving the speed of business. GitLab and Concurrent DevOps collapses cycle times by driving higher efficiency across all stages of the software development lifecycle. For the first time, Product, Development, QA, Security, and Operations teams can work concurrently in a single application. There’s no need to integrate and synchronize tools, or waste time waiting for handoffs. Everyone contributes to a single conversation, instead of managing multiple threads across disparate tools. And only GitLab gives teams complete visibility across the lifecycle with a single, trusted source of data to simplify troubleshooting and drive accountability. All activity is governed by consistent controls, making security and compliance first-class citizens instead of an afterthought. Built on Open Source, GitLab leverages the community contributions of thousands of developers and millions of users to continuously deliver new DevOps innovations. More than 100,000 organizations from startups to global enterprise organizations, including Ticketmaster, Jaguar Land Rover, NASDAQ, Dish Network and Comcast trust GitLab to deliver great software at new speeds.	B2B	Engineering, Product and Design	http://gitlab.com/	Software. Faster.	GitLab is the most comprehensive AI-powered DevSecOps Platform.	Get free trial	2012	San Francisco	W15, PUBLIC, DEVELOPER-TOOLS, DEVSECOPS, OPEN-SOURCE, SAN FRANCISCO

Stack used

I'm using LanceDB to build a MVP fast.

pip install lancedb

Load bi-encoder

db = lancedb.connect("/tmp/db")
model = get_registry().get("sentence-transformers").create(name="BAAI/bge-small-en-v1.5", device="cpu")

Load metadata

class Documents(LanceModel):
    text: str = model.SourceField()
    vector: Vector(model.ndims()) = model.VectorField()

table = db.create_table("pd_table", schema=Documents)

Read data

df = pd.read_csv("data.csv")

Transform data into required format

df["text"] = df[df.columns.to_list()].astype(str).agg(" ".join, axis=1)

docs = [{"text": x} for x in df["text"]]

Add docs to the table

table.add(docs)

Create a index on the column

table.create_fts_index("text")

Reranking

from lancedb.rerankers import CohereReranker

reranker = CohereReranker(api_key=cohere_api_key)

Query

query = "How many repositories does github have?"

result = table.search(query, query_type="hybrid").limit(5).rerank(reranker=reranker).to_list()

Chat

from langchain_cohere import ChatCohere

prompt = f"""
Context is given below:
{context}

Answer the question using the context: {query}
"""

cohere_chat_model = ChatCohere(cohere_api_key=key)
response = cohere_chat_model.invoke(prompt)
print(f"Response: \n {response.content}")