Advanced retrieval: hybrid + reranking

This day is very exam heavy because AWS

loves the words hybrid, filters, and

reranking.

Advanced retrieval, hybrid search plus

re-ranking.

Big idea, one sentence. Good retrieval

is recall first, precision later. Use

hybrid search to catch candidates, then

rerank to put the best answers on top.

One, why vector only? Retrieval is not

enough. Vector search is great at

meaning but bad at exact terms, names,

IDs, codes, dates, legal phrases must

contain constraints. Example, find the

exact policy about NSW stamp duty 2023.

A pure vector search may return similar

tax docs, older years, other states.

That's where hybrid search comes in.

Knack 2 hybrid search keyword vector.

what it means. You combine keyword

search, BM25, text match, vector

similarity, semantic meaning, then merge

the results. This is most commonly done

in open search. Why AWS likes hybrid

search? Because real questions usually

contain intent, how do I, and

constraints in NSW after 2020 policy

123. Hybrid search respects both simple

mental model. Keyword search does it

literally mention this? Vector search

equals is it about the same thing?

Hybrid search asks both.

Nasher 3 metadata filters non-negotiable

in exams. Filters are not search they

are rules. Examples tenant ID equals X

region equals NSW year= 2021. Document

type is policy. Filters reduce noise

enforce security. Improve precision.

Reduce ranker load. Exam signal. If the

question mentions permissions, tenants,

time ranges, regions, and the answer

doesn't include metadata filters, it's

incomplete. Her four, topic

segmentation, quiet but important, what

it is. Instead of one giant index, you

split content by topic, or route queries

to topic specific indexes. Examples: HR

versus legal versus finance, product

docs versus support tickets. Why this

matters? Smaller search space, higher

relevance, cheaper retrieval. Exam

signal, reduce noise, improve relevance,

domain specific search, topic

segmentation.

Number five, reerrankers, precision

layer. This is where many people get

confused. What a re-ranker does. Takes

the top end results from retrieval,

scores them more carefully, reorders

them. It does not search the whole

corpus. Where re-rankers sit important.

You can see the code in our conversation

history. Rankers sit after retrieval,

never before. Number six, what

re-rankers fix. Exam gold re-rankers fix

almost right but wrong order subtle

intent mismatches long context relevance

query nuance they are especially good at

legal text policy docs multicondition

questions they do not fix missing

documents bad filters wrong corpus

garbage in garbage still out number

seven AWS static plus one advanced

retrieval edition static retrieval

strategy hybrid filters topic routing

re-ranker choice top K values plus one

user query the retrieval system is fixed

each query flows through it that's

static plus one again number eight how

AWS expects you to design this end to

end typical enterprise flow one user

query arrives two apply metadata filters

security tenant date run hybrid search

four take top 20 to 50 results five run

re-ranker six top three to five to the

LLM this minimizes hallucinations token

cost irrelevant content text classic

exam traps. Watch closely. Use reranker

instead of vector search. Reranker

searches the whole index. Hybrid search

doesn't need filters. Vector search

replaces keyword search. AWS wants

layers, not replacements. One memory

story. Don't forget this. The courtroom.

Keyword search. Keyword lookup in law

books. Vector search. Understanding

legal meaning. Hybrid search. Use both.

Filters. Jurisdiction and date rules.

Ranker. Senior judge reordering

arguments. LLM equals lawyer explaining

the verdict. The judge doesn't read

every book. They only review the best

arguments. Exam compression rules.

Memorize. Recall first. Hybrid search.

Precision later. Rank. Rules first.

Metadata filters. Smaller scope. Topic

segmentation. If an answer jumps

straight from QLM, it's wrong. What AWS

is really testing. They want to see if

you understand that retrieval is a

pipeline, not a query. Systems that

retrieve well hallucinate less, even

with average models. I'll walk through

one realistic enterprise example step by

step and show where hybrid search

filters, topic segmentation, and

reranking each earn their keep. Real

example, advanced retrieval, hybrid plus

re-ranking. Scenario, legal and policy

assistant for an Australian company. The

company has 500,000 documents, HR

policies, legal contracts, tax and

compliance docs stored as chunks with

embeddings indexed in open search. Users

ask questions like open quote, "What is

the NSW stamp duty exemption for first

home buyers after 2023?" Close quote.

This question has semantic intent

exemptions first home buyers hard

constraints NSW after 2023 exact terms

stamp duty first home buyer vector only

search will struggle keyword only search

will miss nuance so we use layers

step one topic segmentation reduce the

search space before searching the system

routes the query to a topic index legal

tax index not HR not product docs how

simple classifier or rules or comprehend

topic detection. Why this matters?

Instead of 500k docs, maybe 60k, less

noise, faster retrieval, cheaper

reranking. Exam signal, reduce noise,

less domain specific search, topic

segmentation.

Step two, metadata filters, rules before

relevance. Now we apply filters that are

not optional. You can see the code in

our conversation history. These are hard

gates, not ranking signals. Why filters

come early? Enforce security. Enforce

correctness. Prevent irrelevant

documents from ever being seen.

Exams. If permissions, region, or time

matter, filters come first.

Step three, hybrid retrieval recall

layer.

Now, the system runs hybrid search.

Keyword BM25 stamp duty first home buyer

exemption. Vector similarity. Semantic

meaning of the question. Captures

phrasing like concessions, property

transfer tax, eligible buyers. Open

search merges both result sets. Output

top 50 candidate chunks. This step

maximizes recall. Did we retrieve

everything that could be relevant?

Step four, reranker precision layer. Now

comes the reranker model. What it sees

the original user query, the top 50

candidate chunks. What it does scores

each chunk for fine grained relevance.

understands nuance like exemptions

versus reductions, current versus

historical rules, eligibility versus

definition. What it does not do, it does

not search the whole index. It does not

replace filters. It does not fix missing

documents.

Output top five chunks reordered by true

relevance. Exam signal improved

precision, better ranking, nuanced

relevance. Reranker. Step five, LLM

grounding only. Now only the top three

to five re-ranked chunks are sent to the

LLM. Why? Lower token cost, less

hallucination, stronger grounding. The

LLM now answers for NSW first home

buyers after 2023 stamp duty exemptions

apply if

what happens if you remove pieces exam

traps no hybrid search misses exact

legal terms misses older but still

relevant policies

no metadata filters other states leak in

wrong years appear security breach risk

no reranker almost right answers appear

first long policies outrank precise ones

re-ranker without good retrieval

Perfect ranking of the wrong documents

where each component lives. Clear mental

map. You can see the code in our

conversation history.

This ordering matters. AWS exams care

about order. Stash statics one real

world framing. Static hybrid search

configuration filter schema topic

segmentation rules ranker choice top kun

user query. You don't redesign retrieval

per query. You design it once, then feed

queries through it. One memory story.

Lock it in. The legal research team.

Topic segmentation. Assign the right

department. Filters jurisdiction and

year rules. Hybrid search. Junior

lawyers gather cases. Reranker. Senior

lawyer orders relevance. LLM. Junior

explains the conclusion. The senior

lawyer never reads everything. They only

review the best candidates. Ultrashort

exam cheat sheet. Need exact terms plus

meaning. Hybrid. Need security time

tenant filters. Need better ordering

reranker. Need less noise topic

segmentation. Reranker always comes

after retrieval.