---
title: Search that learned what users meant
subtitle: Hybrid search with Learning-to-Rank
re-ranking for a B2B
sales-intelligence product
slug: enterprise-search-ltr
hook: "Keyword search found the words. Vectors
found the meaning. A ranking
model learned which one the user meant: ~25%
better search relevance."
role: Data Scientist
employer: xiQ Inc. (Lahore)
client: xiQ's own B2B sales-intelligence
product (SalesApp and Workbench)
period: Jul 2020 – Dec 2022
tracks:
- data-science
- agentic-genai
tags:
- information retrieval
- hybrid search
- Learning-to-Rank
- Elasticsearch
- vector search
- NLP
metrics:
- value: ~25%
label: Better search relevance metrics,
with Elasticsearch, vector search and
Learning-to-Rank
when: Jul 2020 – Dec 2022
- value: ~60%
label: Less manual content processing,
after automating the curation pipeline
when: Jul 2020 – Dec 2022
---
## Context
xiQ builds sales intelligence: who to call,
what they care about, and what to say. Its
products, SalesApp and Workbench, run on
search. Sales reps look up companies,
people and topics, and they don't type
like librarians.
From July 2020 to December 2022 I was a data
scientist there, working on search, NLP
and recommendations.
## The problem
Each kind of search fails in its own way.
- **Keyword search** (BM25) is precise and
literal. It's great for a company name and
bad at "companies cutting cloud spend".
- **Vector search** understands meaning, but
it fumbles exact names, fresh terms and
acronyms.
- **Users need both**, in one ranked list, in
the order they'd have picked themselves.
## Constraints
- **Interactive latency.** Search sits in the
middle of a rep's workflow; slow is the
same as broken.
- **Mixed queries.** Names, topics and
half-remembered phrases, often in the same
session.
- **Domain language.** Sales and industry
jargon that general-purpose text models
handle badly.
- **2020–2022 tooling.** No LLM re-rankers and
no off-the-shelf hybrid search. You built
the pieces yourself.
## What I built
**NLP-driven search modules for SalesApp and
Workbench.**
- **Keyword retrieval** on Elasticsearch, with
Solr in the stack, for exact and lexical
matches.
- **Vector search** over embeddings, for
meaning.
- **Learning-to-Rank re-ranking** on top of
the merged candidates, so a learned model,
not a hand-tuned formula, decides the
final order.
**Next to search, on the same platform:**
- a content recommendation engine, plus
NLP-based DISC personality analysis and
intent data, to help sales teams target
the right leads;
- an automated content-curation pipeline that
cut manual content processing by ~60%.
## How it fits together
Two retrievers, one ranker.
A query is normalized and checked for
entities. It then goes to two retrievers
in parallel: keyword retrieval for exact
matches, and vector search for meaning.
Their candidates are merged, and a
Learning-to-Rank model re-ranks the merged
list. Relevance judgments train the
ranker. Offline evaluation on those
judgments decides whether a new ranker
ships.
## Stack
- **Search:** Elasticsearch, Apache Solr,
vector search, Learning-to-Rank.
- **NLP and ML:** Python, spaCy, NLTK, Gensim,
scikit-learn.
- **Data and services:** PostgreSQL, MongoDB,
Neo4j, Django, social-media APIs.
- **Cloud:** AWS (S3, EC2, ECS).
## Results
- **~25% better search relevance metrics** in
SalesApp and Workbench.
- **~60% less manual content processing**,
from the curation pipeline.
- **Top Performer of the Year 2021–22** at
xiQ.
## What I'd do differently
This was 2020–2022. Looking back with 2026
eyes:
- **Publish the metric with the number.**
"~25% better relevance" is the honest
figure I have. Today I'd pin every
relevance claim to a named metric (nDCG@10
or recall@k) on a frozen judgment set, and
keep that set in version control.
- **Fuse before you learn.** Reciprocal rank
fusion of the two retrievers is a strong,
cheap baseline. I'd ship it first, and
make the learned ranker beat it.
- **Add a cross-encoder re-ranker and
LLM-assisted judgments.** Both were
expensive or immature then, and both are
standard now. LLM-labeled judgments,
spot-checked by humans, make the judgment
set far cheaper to grow.
- **Give search an eval gate.** A regression
suite of queries that must keep working,
run on every change, the way model code
gets tests.
## What I left out
Not included, on purpose:
- xiQ's corpus, schemas, ranking features and
model internals;
- xiQ's customers and their data;
- the relevance metric's exact definition and
baseline, which are internal to xiQ.
## How to verify
- **LinkedIn project** "Enterprise search with
Learning-to-Rank" (xiQ, Jul 2020 – Dec
2022): hybrid search with Learning-to-Rank
re-ranking and ~25% better search
relevance.
- **LinkedIn experience:** xiQ, Data
Scientist, Jul 2020 – Dec 2022, including
the ~60% content-curation result.
- **LinkedIn honors:** Top Performer of the
Year 2021–22, xiQ.
- **Employment check:** xiQ Inc., Data
Scientist, Jul 2020 – Dec 2022.
- **In an interview:** a walk-through of the
method, within NDA limits.
- **Code:** walkthroughs available on request.
Case study 08 · xiQ Inc. · Jul 2020 – Dec 2022
Search that learned what users meant
Hybrid search with Learning-to-Rank re-ranking for a B2B sales-intelligence product
Keyword search found the words. Vectors found the meaning. A ranking model learned which one the user meant: ~25% better search relevance.
✓ Compiled from this page's own markdown.
Context
xiQ builds sales intelligence: who to call, what they care about, and what to say. Its products, SalesApp and Workbench, run on search. Sales reps look up companies, people and topics, and they don't type like librarians.
From July 2020 to December 2022 I was a data scientist there, working on search, NLP and recommendations.
The problem
Each kind of search fails in its own way.
- Keyword search (BM25) is precise and literal. It's great for a company name and bad at "companies cutting cloud spend".
- Vector search understands meaning, but it fumbles exact names, fresh terms and acronyms.
- Users need both, in one ranked list, in the order they'd have picked themselves.
Constraints
- Interactive latency. Search sits in the middle of a rep's workflow; slow is the same as broken.
- Mixed queries. Names, topics and half-remembered phrases, often in the same session.
- Domain language. Sales and industry jargon that general-purpose text models handle badly.
- 2020–2022 tooling. No LLM re-rankers and no off-the-shelf hybrid search. You built the pieces yourself.
What I built
NLP-driven search modules for SalesApp and Workbench.
- Keyword retrieval on Elasticsearch, with Solr in the stack, for exact and lexical matches.
- Vector search over embeddings, for meaning.
- Learning-to-Rank re-ranking on top of the merged candidates, so a learned model, not a hand-tuned formula, decides the final order.
Next to search, on the same platform:
- a content recommendation engine, plus NLP-based DISC personality analysis and intent data, to help sales teams target the right leads;
- an automated content-curation pipeline that cut manual content processing by ~60%.
How it fits together
Two retrievers, one ranker.
A query is normalized and checked for entities. It then goes to two retrievers in parallel: keyword retrieval for exact matches, and vector search for meaning. Their candidates are merged, and a Learning-to-Rank model re-ranks the merged list. Relevance judgments train the ranker. Offline evaluation on those judgments decides whether a new ranker ships.
Stack
- Search: Elasticsearch, Apache Solr, vector search, Learning-to-Rank.
- NLP and ML: Python, spaCy, NLTK, Gensim, scikit-learn.
- Data and services: PostgreSQL, MongoDB, Neo4j, Django, social-media APIs.
- Cloud: AWS (S3, EC2, ECS).
Results
- ~25% better search relevance metrics in SalesApp and Workbench.
- ~60% less manual content processing, from the curation pipeline.
- Top Performer of the Year 2021–22 at xiQ.
What I'd do differently
This was 2020–2022. Looking back with 2026 eyes:
- Publish the metric with the number. "~25% better relevance" is the honest figure I have. Today I'd pin every relevance claim to a named metric (nDCG@10 or recall@k) on a frozen judgment set, and keep that set in version control.
- Fuse before you learn. Reciprocal rank fusion of the two retrievers is a strong, cheap baseline. I'd ship it first, and make the learned ranker beat it.
- Add a cross-encoder re-ranker and LLM-assisted judgments. Both were expensive or immature then, and both are standard now. LLM-labeled judgments, spot-checked by humans, make the judgment set far cheaper to grow.
- Give search an eval gate. A regression suite of queries that must keep working, run on every change, the way model code gets tests.
What I left out
Not included, on purpose:
- xiQ's corpus, schemas, ranking features and model internals;
- xiQ's customers and their data;
- the relevance metric's exact definition and baseline, which are internal to xiQ.
How to verify
- LinkedIn project "Enterprise search with Learning-to-Rank" (xiQ, Jul 2020 – Dec 2022): hybrid search with Learning-to-Rank re-ranking and ~25% better search relevance.
- LinkedIn experience: xiQ, Data Scientist, Jul 2020 – Dec 2022, including the ~60% content-curation result.
- LinkedIn honors: Top Performer of the Year 2021–22, xiQ.
- Employment check: xiQ Inc., Data Scientist, Jul 2020 – Dec 2022.
- In an interview: a walk-through of the method, within NDA limits.
- Code: walkthroughs available on request.