Beyond Vector Search: Building a Deterministic 3-Tiered Graph-RAG System

On this article, you’ll learn to construct a deterministic, multi-tier retrieval-augmented technology system utilizing information graphs and vector databases.

Subjects we’ll cowl embody:

Designing a three-tier retrieval hierarchy for factual accuracy.
Implementing a light-weight information graph.
Utilizing prompt-enforced guidelines to resolve retrieval conflicts deterministically.

Past Vector Search: Constructing a Deterministic 3-Tiered Graph-RAG System
Picture by Editor

Introduction: The Limits of Vector RAG

Vector databases have lengthy since turn out to be the cornerstone of contemporary retrieval augmented technology (RAG) pipelines, excelling at retrieving long-form textual content primarily based on semantic similarity. Nonetheless, vector databases are notoriously “lossy” on the subject of atomic details, numbers, and strict entity relationships. A typical vector RAG system would possibly simply confuse which crew a basketball participant presently performs for, for instance, just because a number of groups seem close to the participant’s identify in latent house. To unravel this, we want a multi-index, federated structure.

On this tutorial, we’ll introduce such an structure, utilizing a quad retailer backend to implement a information graph for atomic details, backed by a vector database for long-tail, fuzzy context.

However right here is the twist: as a substitute of counting on complicated algorithmic routing to choose the suitable database, we’ll question all databases, dump the outcomes into the context window, and use prompt-enforced fusion guidelines to pressure the language mannequin (LM) to deterministically resolve conflicts. The purpose is to aim to remove relationship hallucinations and construct absolute deterministic predictability the place it issues most: atomic details.

Structure Overview: The three-Tiered Hierarchy

Our pipeline enforces strict information hierarchy utilizing three retrieval tiers:

Precedence 1 (absolute graph details): A easy Python QuadStore information graph containing verified, immutable floor truths structured in Topic-Predicate-Object plus Context (SPOC) format.
Precedence 2 (statistical graph information): A secondary QuadStore containing aggregated statistics or historic information. This tier is topic to Precedence 1 override in case of conflicts (e.g. a Precedence 1 present crew truth overrides a Precedence 2 historic crew statistic).
Precedence 3 (vector paperwork): A typical dense vector DB (ChromaDB) for common textual content paperwork, solely used as a fallback if the information graphs lack the reply.

Surroundings & Conditions Setup

To observe alongside, you will want an surroundings working Python, a neighborhood LM infrastructure and served mannequin (we use Ollama with llama3.2), and the next core libraries:

chromadb: For the vector database tier
spaCy: For named entity recognition (NER) to question the graphs
requests: To work together with our native LM inference endpoint
QuadStore: For the information graph tier (see QuadStore repository)

# Set up required libraries
pip set up chromadb spacy requests

# Obtain the spaCy English mannequin
python -m spacy obtain en_core_web_sm

# Set up required libraries

pip set up chromadb spacy requests

# Obtain the spaCy English mannequin

python –m spacy obtain en_core_web_sm

You possibly can manually obtain the straightforward Python QuadStore implementation from the QuadStore repository and place it someplace in your native file system to import as a module.

⚠️ Notice: The total mission code implementation is accessible on this GitHub repository.

With these stipulations dealt with, let’s dive into the implementation.

Step 1: Constructing a Light-weight QuadStore (The Graph)

To implement Precedence 1 and Precedence 2 information, we use a customized light-weight in-memory information graph referred to as a quad retailer. This data graph shifts away from semantic embeddings towards a strict node-edge-node schema identified internally as a SPOC (Topic-Predicate-Object plus Context).

This QuadStore module operates as a highly-indexed storage engine. Underneath the hood, it maps all strings into integer IDs to stop reminiscence bloat, whereas conserving a four-way dictionary index (spoc, pocs, ocsp, cspo) to allow constant-time lookups throughout any dimension. Whereas we received’t dive into the main points of the inner construction of the engine right here, using the API in our RAG script is extremely easy.

Why use this straightforward implementation as a substitute of a extra sturdy graph database like Neo4j or ArangoDB? Simplicity and velocity. This implementation is extremely light-weight and quick, whereas having the extra advantage of being simple to know. That is all that’s wanted for this particular use case with out having to be taught a posh graph database API.

There are actually solely a few QuadStore strategies it is advisable perceive:

add(topic, predicate, object, context): Provides a brand new truth to the information graph
question(topic, predicate, object, context): Queries the information graph for details that match the given topic, predicate, object, and context

Let’s initialize the QuadStore appearing as our Precedence 1 absolute fact mannequin:

from quadstore import QuadStore

# Initialize details quadstore
facts_qs = QuadStore()

# Natively add details (Topic, Predicate, Object, Context)
facts_qs.add(“LeBron James”, “likes”, “coconut milk”, “NBA_trivia”)
facts_qs.add(“LeBron James”, “played_for”, “Ottawa Beavers”, “NBA_2023_regular_season”)
facts_qs.add(“Ottawa Beavers”, “obtained”, “LeBron James”, “2020_expansion_draft”)
facts_qs.add(“Ottawa Beavers”, “based_in”, “downtown Ottawa”, “NBA_trivia”)
facts_qs.add(“Kevin Durant”, “is”, “an individual”, “NBA_trivia”)
facts_qs.add(“Ottawa Beavers”, “had”, “worst first 12 months of any enlargement crew in NBA historical past”, “NBA_trivia”)
facts_qs.add(“LeBron James”, “average_mpg”, “12.0”, “NBA_2023_regular_season”)

from quadstore import QuadStore

# Initialize details quadstore

facts_qs = QuadStore()

# Natively add details (Topic, Predicate, Object, Context)

facts_qs.add(“LeBron James”, “likes”, “coconut milk”, “NBA_trivia”)

facts_qs.add(“LeBron James”, “played_for”, “Ottawa Beavers”, “NBA_2023_regular_season”)

facts_qs.add(“Ottawa Beavers”, “obtained”, “LeBron James”, “2020_expansion_draft”)

facts_qs.add(“Ottawa Beavers”, “based_in”, “downtown Ottawa”, “NBA_trivia”)

facts_qs.add(“Kevin Durant”, “is”, “an individual”, “NBA_trivia”)

facts_qs.add(“Ottawa Beavers”, “had”, “worst first 12 months of any enlargement crew in NBA historical past”, “NBA_trivia”)

facts_qs.add(“LeBron James”, “average_mpg”, “12.0”, “NBA_2023_regular_season”)

As a result of it makes use of the equivalent underlying class, you’ll be able to populate Precedence 2 (which handles broader statistics and numbers) identically or by studying from a previously-prepared JSONLines file. This file was created by working a easy script that learn the 2023 NBA common season stats from a CSV file that was freely-acquired from a basketball stats web site (although I can not recall which one, as I’ve had the info for a number of years at this level), and transformed every row right into a quad. You possibly can obtain the pre-processed NBA 2023 stats file in JSONL format from the mission repository.

Step 2: Integrating the Vector Database

Subsequent, we set up our Precedence 3 layer: the usual dense vector DB. We use ChromaDB to retailer textual content chunks that our inflexible information graphs may need missed.

Right here is how we initialize a persistent assortment and ingest uncooked textual content into it:

import chromadb
from chromadb.config import Settings

# Initialize vector embeddings
chroma_client = chromadb.PersistentClient(
path=”./chroma_db”,
settings=Settings(anonymized_telemetry=False)
)
assortment = chroma_client.get_or_create_collection(identify=”basketball”)

# Our fallback unstructured textual content chunks
doc1 = (
“LeBron injured for the rest of NBA 2023 seasonn”
“LeBron James suffered an ankle damage early within the season, which led to him taking part in far “
“fewer minutes per sport than he has not too long ago averaged in different seasons. The damage obtained a lot “
“worse at the moment, and he’s out for the remainder of the season.”
)
doc2 = (
“Ottawa Beaversn”
“The Ottawa Beavers star participant LeBron James is out for the remainder of the 2023 NBA season, “
“after his ankle damage has worsened. The groups’ abysmal common season report might find yourself “
“being the worst of any crew ever, with solely 6 wins as of now, with solely 4 gmaes left in “
“the common season.”
)

assortment.upsert(
paperwork=[doc1, doc2],
ids=[“doc1”, “doc2”]
)

import chromadb

from chromadb.config import Settings

# Initialize vector embeddings

chroma_client = chromadb.PersistentClient(

path=“./chroma_db”,

settings=Settings(anonymized_telemetry=False)

)

assortment = chroma_client.get_or_create_collection(identify=“basketball”)

# Our fallback unstructured textual content chunks

doc1 = (

“LeBron injured for the rest of NBA 2023 seasonn”

“LeBron James suffered an ankle damage early within the season, which led to him taking part in far “

“fewer minutes per sport than he has not too long ago averaged in different seasons. The damage obtained a lot “

“worse at the moment, and he’s out for the remainder of the season.”

)

doc2 = (

“Ottawa Beaversn”

“The Ottawa Beavers star participant LeBron James is out for the remainder of the 2023 NBA season, “

“after his ankle damage has worsened. The groups’ abysmal common season report might find yourself “

“being the worst of any crew ever, with solely 6 wins as of now, with solely 4 gmaes left in “

“the common season.”

)

assortment.upsert(

paperwork=[doc1, doc2],

ids=[“doc1”, “doc2”]

)

Step 3: Entity Extraction & International Retrieval

How can we question deterministic graphs and semantic vectors concurrently? We bridge the hole utilizing NER by way of spaCy.

First, we extract entities in fixed time from the consumer’s immediate (e.g. “LeBron James” and “Ottawa Beavers”). Then, we fireplace off parallel queries to each QuadStores utilizing the entities as strict lookups, whereas querying ChromaDB utilizing string similarity over the immediate content material.

import spacy

# Load our NLP mannequin
nlp = spacy.load(“en_core_web_sm”)

def extract_entities(textual content):
“””
Extract entities from the given textual content utilizing spaCy. Utilizing set eliminates duplicates.
“””
doc = nlp(textual content)
return record(set([ent.text for ent in doc.ents]))

def get_facts(qs, entities):
“””
Retrieve details for a listing of entities from the QuadStore (querying topics and objects).
“””
details = []
for entity in entities:
subject_facts = qs.question(topic=entity)
object_facts = qs.question(object=entity)
details.lengthen(subject_facts + object_facts)
# Deduplicate details and return
return record(set(tuple(truth) for truth in details))

import spacy

# Load our NLP mannequin

nlp = spacy.load(“en_core_web_sm”)

def extract_entities(textual content):

“”“

Extract entities from the given textual content utilizing spaCy. Utilizing set eliminates duplicates.

““”

doc = nlp(textual content)

return record(set([ent.text for ent in doc.ents]))

def get_facts(qs, entities):

“”“

Retrieve details for a listing of entities from the QuadStore (querying topics and objects).

““”

details = []

for entity in entities:

subject_facts = qs.question(topic=entity)

object_facts = qs.question(object=entity)

details.lengthen(subject_facts + object_facts)

# Deduplicate details and return

return record(set(tuple(truth) for truth in details))

We now have all of the retrieved context separated into three distinct streams (facts_p1, facts_p2, and vec_info).

Step 4: Immediate-Enforced Battle Decision

Usually, complicated algorithmic battle decision (like Reciprocal Rank Fusion) fails when resolving granular details towards broad textual content. Right here we take a radically easier method that, as a sensible matter, additionally appears to work nicely: we embed the “adjudicator” ruleset immediately into the system immediate.

By assembling the information into explicitly labeled [PRIORITY 1], [PRIORITY 2], and [PRIORITY 3] blocks, we instruct the language mannequin to observe specific logic when outputting its response.

Right here is the system immediate in its entirety:

def create_system_prompt(details, stats, data):
# Format graph details into easy declarative sentences for language mannequin comprehension
formatted_facts = “n”.be a part of([f”In {q[3]}, {q[0]} {str(q[1]).exchange(‘_’, ‘ ‘)} {q[2]}.” if len(q) >= 4 else str(q) for q in details])
formatted_stats = “n”.be a part of([f”In {q[3]}, {q[0]} {str(q[1]).exchange(‘_’, ‘ ‘)} {q[2]}.” if len(q) >= 4 else str(q) for q in stats])

# Convert retrieved data dict to a string of textual content paperwork
retrieved_context = “”
if data and ‘paperwork’ in data and data[‘documents’]:
retrieved_context = ” “.be a part of(data[‘documents’][0])

return f”””You’re a strict data-retrieval AI. Your ONLY information comes from the textual content offered beneath. You could fully ignore your inside coaching weights.

PRIORITY RULES (strict):
1. If Precedence 1 (Details) accommodates a direct reply, use ONLY that reply. Don’t complement, qualify, or cross-reference with Precedence 2 or Vector information.
2. Precedence 2 information makes use of abbreviations and will seem to contradict P1 — it’s supplementary background solely. By no means deal with P2 crew abbreviations as authoritative crew names if P1 states a crew.
3. Solely use P2 if P1 has no related reply on the precise attribute requested.
4. If Precedence 3 (Vector Chunks) supplies any extra related data, use your judgment as as to whether or to not embody it within the response.
5. If not one of the sections include the reply, you will need to explicitly say “I would not have sufficient data.” Don’t guess or hallucinate.

Your output **MUST** observe these guidelines:
– Present solely the only authoritative reply primarily based on the precedence guidelines.
– Don’t current a number of conflicting solutions.
– Make no point out of the supply of this information.
– Phrase this within the type of a sentence or a number of sentences, as is suitable.

—
[PRIORITY 1 – ABSOLUTE GRAPH FACTS]
{formatted_facts}

[Priority 2: Background Statistics (team abbreviations here are NOT authoritative — defer to Priority 1 for factual claims)]
{formatted_stats}

[PRIORITY 3 – VECTOR DOCUMENTS]
{retrieved_context}
—
“””

def create_system_prompt(details, stats, data):

# Format graph details into easy declarative sentences for language mannequin comprehension

formatted_facts = “n”.be a part of([f“In {q[3]}, {q[0]} {str(q[1]).exchange(‘_’, ‘ ‘)} {q[2]}.” if len(q) >= 4 else str(q) for q in details])

formatted_stats = “n”.be a part of([f“In {q[3]}, {q[0]} {str(q[1]).exchange(‘_’, ‘ ‘)} {q[2]}.” if len(q) >= 4 else str(q) for q in stats])

# Convert retrieved data dict to a string of textual content paperwork

retrieved_context = “”

if data and ‘paperwork’ in data and data[‘documents’]:

retrieved_context = ” “.be a part of(data[‘documents’][0])

return f“”“You’re a strict data-retrieval AI. Your ONLY information comes from the textual content offered beneath. You could fully ignore your inside coaching weights.

PRIORITY RULES (strict):

1. If Precedence 1 (Details) accommodates a direct reply, use ONLY that reply. Don’t complement, qualify, or cross-reference with Precedence 2 or Vector information.

2. Precedence 2 information makes use of abbreviations and will seem to contradict P1 — it’s supplementary background solely. By no means deal with P2 crew abbreviations as authoritative crew names if P1 states a crew.

3. Solely use P2 if P1 has no related reply on the precise attribute requested.

4. If Precedence 3 (Vector Chunks) supplies any extra related data, use your judgment as as to whether or to not embody it within the response.

5. If not one of the sections include the reply, you will need to explicitly say “I do not have sufficient data.” Don’t guess or hallucinate.

Your output **MUST** observe these guidelines:

– Present solely the only authoritative reply primarily based on the precedence guidelines.

– Don’t current a number of conflicting solutions.

– Make no point out of the supply of this information.

– Phrase this within the type of a sentence or a number of sentences, as is suitable.

—

[PRIORITY 1 – ABSOLUTE GRAPH FACTS]

{formatted_facts}

[Priority 2: Background Statistics (team abbreviations here are NOT authoritative — defer to Priority 1 for factual claims)]

{formatted_stats}

[PRIORITY 3 – VECTOR DOCUMENTS]

{retrieved_context}

—

““”

Far completely different than “… and don’t make any errors” prompts which might be little greater than finger-crossing and wishing for no hallucinations, on this case we current the LM with floor fact atomic details, attainable conflicting “less-fresh” details, and semantically-similar vector search outcomes, together with an specific hierarchy for figuring out which set of knowledge is appropriate when conflicts are encountered. Is it foolproof? No, in fact not, however it’s a distinct method worthy of consideration and addition to the hallucination-combatting toolkit.

Don’t neglect that you’ll find the remainder of the code for this mission right here.

Step 5: Tying it All Collectively & Testing

To wrap all the things up, the primary execution thread of our RAG system calls the native Llama occasion by way of the REST API, handing it the structured system immediate above alongside the consumer’s base query.

When run within the terminal, the system isolates our three precedence tiers, processes the entities, and queries the LM deterministically.

Question 1: Factual Retrieval with the QuadStore

When querying an absolute truth like “Who’s the star participant of Ottawa Beavers crew?”, the system depends totally on Precedence 1 details.

LeBron performs for Ottawa Beavers

As a result of Precedence 1, on this case, explicitly states “Ottawa Beavers obtained LeBron James”, the immediate instructs the LM by no means to complement this with the vector paperwork or statistical abbreviations, thus aiming to remove the normal RAG relationship hallucination. The supporting vector database paperwork assist this declare as nicely, with articles about LeBron and his tenure with the Ottawa NBA crew. Evaluate this with an LM immediate that dumps conflicting semantic search outcomes right into a mannequin and asks it, generically, to find out which is true.

Question 2: Extra Factual Retrieval

The Ottawa beavers, you say? I’m unfamiliar with them. I assume they play out of Ottawa, however the place, precisely, within the metropolis are they primarily based? Precedence 1 details can inform us. Take into accout we’re preventing towards what the mannequin itself already is aware of (the Beavers aren’t an precise NBA crew) in addition to the NBA common stats dataset (which lists nothing in regards to the Ottawa Beavers in any way).

The Ottawa Beavers dwelling

Question 3: Coping with Battle

When querying an attribute in each absolutely the details graph and the final stats graph, equivalent to “What was LeBron James’ common MPG within the 2023 NBA season?”, the mannequin depends on the Precedence 1 stage information over the prevailing Precedence 2 stats information.

LeBron MPG Question Output

Question 4: Stitching Collectively a Strong Response

What occurs after we ask an unstructured query like “What damage did the Ottawa Beavers star damage undergo in the course of the 2023 season?” First, the mannequin must know who the Ottawa Beavers star participant is, after which decide what their damage was. That is completed with a mixture of Precedence 1 and Precedence 3 information. The LM merges this easily right into a closing response.

LeBron Damage Question Output

Question 5: One other Strong Response

Right here’s one other instance of sewing collectively a coherent and correct response from multi-level information. “What number of wins did the crew that LeBron James play for have when he left the season?”

LeBron Damage Question #2 Output

Let’s not neglect that for all of those queries, the mannequin should ignore the truth that conflicting (and inaccurate!) information exists within the Precedence 2 stats graph suggesting (once more, wrongly!) that LeBron James performed for the LA Lakers in 2023. And let’s additionally not neglect that we’re utilizing a easy language mannequin with solely 3 billion parameters (llama3.2:3b).

Conclusion & Commerce-offs

By splitting your retrieval sources into distinct authoritative layers — and dictating precise decision guidelines by way of immediate engineering — the hope is that you simply drastically scale back factual hallucinations, or competitors between in any other case equally-true items of knowledge.

Benefits of this method embody:

Predictability: 100% deterministic predictability for essential details saved in Precedence 1 (purpose)
Explainability: If required, you’ll be able to pressure the LM to output its [REASONING] chain to validate why Precedence 1 overrode the remaining
Simplicity: No want to coach customized retrieval routers

Commerce-offs of this method embody:

Token Overhead: Dumping all three databases into the preliminary context window consumes considerably extra tokens than typical algorithm-filtered retrieval
Mannequin Reliance: This method requires a extremely instruction-compliant LM to keep away from falling again into latent training-weight conduct

For environments by which excessive precision and low tolerance for errors are necessary, deploying a multi-tiered factual hierarchy alongside your vector database will be the differentiator between prototype and manufacturing.