safe_store: The Local Multi-Modal Vector, Graph & Semantic Engine

safe_store is an ultra-fast, local, and sovereign knowledge engine for Python. It transforms unstructured documents (PDF, DOCX, HTML, Markdown, Code) and structured datasets (CSV, Excel XLSX, SQLite) into an interconnected, queryable knowledge base combining:

🧠 Dense Semantic Vector Search: Embeddings powered by Sentence-Transformers, Ollama, OpenAI, Cohere, Lollms, or TF-IDF.
⚡ Sparse Lexical Search (BM25): Native SQLite FTS5 full-text indexing for exact technical identifiers, part numbers, and error codes.
📖 Full Document & Context Window Retrieval: Query entire documents aggregated from chunk hits, retrieve surrounding chunk neighborhoods with window expansion, or paginate through document content.
🕸️ Dynamic & Ontology Knowledge Graph: Open-ended concept/entity extraction from text files or strict TBox/OWL schema mapping, with live chunk extraction reporting.
🔍 W3C SPARQL 1.1 Query & Update Engine: Native TBox/ABox ontology management, declarative tabular mapping, and full SPARQL (SELECT, ASK, CONSTRUCT, DESCRIBE, and INSERT/DELETE DATA updates).
🧠 LLM Cognitive Memory & Thought Reorganization: Episodic memory logging, associative semantic traversal, grounded text chunk evidence linking, and native function-calling tool dispatching.
🔀 Tri-Modal Reciprocal Rank Fusion (RRF): Merges dense similarity, lexical BM25, and symbolic graph traversals into unified results with universal 0–100 relevance grades.
🔍 Database Diagnostics & Introspection (store.info()): Instant inspection of vectorizers, chunking parameters, document chunk counts, ontology schemas, and graph topology metrics.
📊 Semantic Datalake & Point Cloud Engine: 2D/3D PCA & t-SNE projections with persistent SQLite caching, streaming lazy loading (IncrementalPCA), and interactive HTML visualizer exports.
🔐 Zero-Leakage Local Encryption: End-to-end AES-128/HMAC (Fernet) encryption at rest inside a single, portable .db file.

📦 Installation

pip install safe_store

🌟 Core Architecture & Pillars

                               ┌────────────────────────────────────────┐
                               │           User Natural Query           │
                               └──────────────────┬─────────────────────┘
                                                  │
                ┌─────────────────────────────────┼─────────────────────────────────┐
                ▼                                 ▼                                 ▼
    ┌───────────────────────┐         ┌───────────────────────┐         ┌───────────────────────┐
    │  Dense Vector Search  │         │   Sparse BM25 Search  │         │  Symbolic Graph Query │
    │  (Semantic Context)   │         │ (Exact IDs/SKUs/Names)│         │ (TBox/ABox/SPARQL/Hop)│
    └───────────┬───────────┘         └───────────┬───────────┘         └───────────┬───────────┘
                │                                 │                                 │
                │        [Candidate Set 1]        │        [Candidate Set 2]        │ [Candidate Set 3]
                └─────────────────────────────────┼─────────────────────────────────┘
                                                  ▼
                               ┌─────────────────────────────────────┐
                               │  Reciprocal Rank Fusion (RRF / WCS) │
                               │  Score = Σ (w_i / (k + rank_i))     │
                               └──────────────────┬──────────────────┘
                                                  ▼
                               ┌─────────────────────────────────────┐
                               │  Enriched Context + Provenance Lineage│
                               └──────────────────┬──────────────────┘
                                                  ▼
                               ┌─────────────────────────────────────┐
                               │       LLM Response Generation       │
                               └─────────────────────────────────────┘

🚀 Quick Start

0. Instant Database Introspection & Diagnostics (`store.info()`)

You can inspect any database state, vectorizer configuration, per-document chunk counts, ontology schemas, and knowledge graph metrics with a single method call:

import safe_store

store = safe_store.SafeStore("knowledge.db")

# 1. Print formatted diagnostics panel to console
store.info()

# 2. Or retrieve structured dictionary for APIs and dashboards
db_info = store.get_database_info()
print(f"Total Docs: {db_info['documents']['total_documents']}")
for doc in db_info['documents']['list']:
    print(f"  • {doc['document_title']}: {doc['chunk_count']} chunks")

print(f"Knowledge Graph: {db_info['knowledge_graph']['total_nodes']} nodes, {db_info['knowledge_graph']['total_relationships']} edges")

1. Tri-Modal Hybrid Retrieval (Dense Vectors + BM25 Lexical + RRF)

Combining dense embeddings with sparse BM25 guarantees precision for both fuzzy conceptual questions and exact code/identifier queries.

import safe_store

store = safe_store.SafeStore(
    db_path="hybrid_kb.db",
    vectorizer_name="st",
    vectorizer_config={"model": "all-MiniLM-L6-v2"},
    chunk_size=128,
    chunk_overlap=16
)

# Inspect database summary and diagnostics anytime:
store.info()

1.1 Full Document & Neighborhood Context Window Retrieval

When an LLM needs complete document context or the continuous paragraph surrounding a chunk match:

with store:
    # 1. Full Document Retrieval: Discovers matching chunks, aggregates scores on 0-100 grade,
    #    and excludes documents under the relevance threshold (e.g. min_relevance_percent=50.0)
    full_docs = store.query_full_documents(
        query_text="memory leak troubleshooting",
        top_k_docs=1,
        search_mode='hybrid',
        min_relevance_percent=50.0 # Prevents retrieving irrelevant docs
    )
    if full_docs:
        print(f"Top Document: {full_docs[0]['document_title']} (Relevance: {full_docs[0]['relevance_score']:.1f}%)")
        print(f"Full Text:\n{full_docs[0]['full_text']}\n")
    else:
        print("No document exceeded the 50% relevance threshold.")

    # 2. Window Expansion Retrieval: Expands matching chunks by window_before / window_after chunks

1.2 Unstructured File Ingestion & Dynamic Knowledge Graph Construction

safe_store allows you to extract rich knowledge graphs directly from unstructured files (Markdown, PDF, DOCX, Text) with live per-chunk extraction reporting:

from safe_store import SafeStore, GraphStore

store = SafeStore("project_kb.db", vectorizer_name="st")

with store:
    # 1. Ingest unstructured documents
    store.add_document("architecture_notes.md", metadata={"source": "Design Team"})
    store.add_document("incident_report.pdf", metadata={"source": "SRE"})

    # 2. Initialize GraphStore
    # When no ontology is supplied, it operates in dynamic extraction mode (concepts, tools, entities, relations)
    graph = GraphStore(store=store, llm_executor_callback=my_llm_callback)

    # 3. Build graph across all documents with real-time progress & node/edge reporting
    stats = graph.build_graph_for_all_documents()
    print(f"Graph build finished: {stats['nodes_created']} nodes, {stats['relationships_created']} relationships.")

    # 4. Inspect graph statistics
    graph_info = graph.get_graph_info()
    print(f"Nodes by Label: {graph_info['nodes_by_label']}")
    print(f"Edges by Type:  {graph_info['relationships_by_type']}")

    # 2. Window Expansion Retrieval: Expands matching chunks by window_before / window_after chunks
    windows = store.query_document_content_window(
        query_text="ERR-4091 supervisor daemon",
        top_k_hits=1,
        window_before=1,
        window_after=1,
        min_relevance_percent=40.0
    )
    if windows:
        print(f"Stitched Window Text:\n{windows[0]['stitched_window_text']}\n")

    # 3. Document Chunk Pagination: Browse chunks page by page with sequence tracking
    page_data = store.get_document_content_paginated("incident_001", page=1, page_size=5)
    print(f"Page {page_data['page']} of {page_data['total_pages']} (Total Chunks: {page_data['total_chunks']})")
    print(f"Stitched Page Text:\n{page_data['stitched_text']}")

with store: # Index unstructured technical documents store.add_text( unique_id="incident_001", text="Production node crashed due to OOMKilled condition in supervisor daemon. " "Error code ERR-4091 was emitted by telemetry controller.", metadata={"service": "Telemetry", "severity": "Critical"} ) store.add_text( unique_id="manual_001", text="Troubleshooting Guide: When encountering error code ERR-4091, replace the " "memory buffer chip and execute supervisor restart.", metadata={"doc_type": "Runbook"} )

# Hybrid Query: Score-Calibrated Fusion of Dense Semantic Similarity with BM25 Sparse Lexical Score
results = store.hybrid_query(
    query_text="troubleshooting memory failure ERR-4091",
    top_k=2,
    dense_weight=0.5,
    bm25_weight=0.5,
    rrf_k=60,
    min_relevance_percent=40.0 # Standard 0-100 threshold filter
)

for r in results:
    print(f"[{r['file_path']}] (Relevance: {r['relevance_score']:.1f}% | Raw RRF: {r['raw_rrf_score']:.5f})")
    print(f"Content: {r['chunk_text']}\n")


---

### 2. LLM Cognitive Memory & SPARQL 1.1 Reorganization

Empower LLM agents to reorganize thoughts, record episodic memory events, and traverse associative concept graphs grounded in physical document chunks:

```python
from safe_store import SafeStore, GraphStore

store = SafeStore(db_path="agent_memory.db", vectorizer_name="st")
graph = GraphStore(store=store)

# 1. LLM Reorganizes Knowledge Graph via SPARQL 1.1 UPDATE
graph.execute_sparql_update("""
PREFIX ont: <http://example.org/ontology/>
PREFIX ex: <http://example.org/>
INSERT DATA {
    ex:Alice a ont:Architect ;
             ont:name "Alice Smith" ;
             ont:leadsProject ex:ProjectPhoenix .
    ex:ProjectPhoenix a ont:Project ;
                      ont:status "Active" .
}
""")

# 2. Record an Episodic Event with Chunk Grounding
episode_id = graph.memory.record_episode(
    title="Architecture Design Review",
    description="Alice presented the decentralized ledger protocol for Project Phoenix.",
    participants=["Alice Smith"],
    outcome="Approved",
    source_chunk_ids=[1] # Grounded in chunk #1
)

# 3. Associative Recall: Traverse Semantic Neighborhoods & Evidence
memory_view = graph.memory.recall_associative("Alice Smith", max_hops=2)
print("Associated Entities:", [e['properties']['name'] for e in memory_view['associated_entities']])
print("Source Chunk Evidence:", memory_view['grounded_chunks'][0]['chunk_text'])

# 4. Expose Standard Function-Calling Tools to LLM Agents
llm_tools = graph.get_tool_definitions()
# Pass llm_tools directly to OpenAI, Anthropic, Ollama, or Lollms tool definitions!

3. W3C SPARQL 1.1 Knowledge Graph Engine

safe_store provides a full, standards-compliant SPARQL 1.1 engine supporting SELECT, ASK, CONSTRUCT, and DESCRIBE queries across multi-hop relational graphs.

from safe_store import SafeStore, GraphStore

store = SafeStore(db_path="enterprise_kg.db", vectorizer_name="st")
graph = GraphStore(store=store)

# Create Graph Entities and Relationships
alice_id = graph.add_node("Person", {"name": "Alice Smith", "role": "Lead Architect"})
bob_id = graph.add_node("Person", {"name": "Bob Jones", "role": "Data Scientist"})
acme_id = graph.add_node("Company", {"name": "Acme Robotics", "industry": "AI"})
paris_id = graph.add_node("City", {"name": "Paris", "country": "France"})

graph.add_relationship(alice_id, acme_id, "worksFor", {"since": 2021})
graph.add_relationship(bob_id, acme_id, "worksFor", {"since": 2023})
graph.add_relationship(acme_id, paris_id, "locatedIn")
graph.add_relationship(alice_id, bob_id, "collaboratesWith")

# 1. SPARQL SELECT: Multi-Hop Relational Traversal
sparql_select = """
PREFIX ex: <http://example.org/>
PREFIX ont: <http://example.org/ontology/>
SELECT ?personName ?cityName WHERE {
    ?person ont:worksFor ?company ;
            ont:hasName ?personName .
    ?company ont:locatedIn ?city .
    ?city ont:hasName ?cityName .
}
"""
results = graph.query_sparql(sparql_select)
for b in results["results"]["bindings"]:
    print(f"Person: {b['personName']['value']} works in City: {b['cityName']['value']}")

# 2. SPARQL ASK: Boolean Verification
sparql_ask = """
PREFIX ont: <http://example.org/ontology/>
ASK {
    ?person ont:worksFor ?company .
    ?company ont:hasName "Acme Robotics" .
}
"""
is_valid = graph.query_sparql(sparql_ask)
print(f"Acme Robotics employs personnel: {is_valid['boolean']}")

# 3. SPARQL CONSTRUCT: Subgraph Transformation
sparql_construct = """
PREFIX ont: <http://example.org/ontology/>
PREFIX foaf: <http://xmlns.com/foaf/0.1/>
CONSTRUCT {
    ?person foaf:workplaceHomepage ?company .
}
WHERE {
    ?person ont:worksFor ?company .
}
"""
subgraph = graph.query_sparql(sparql_construct)
for triple in subgraph["triples"]:
    print(f"Constructed: {triple['subject']['value']} -> {triple['predicate']['value']} -> {triple['object']['value']}")

3. TBox (Ontology) & Declarative Tabular-to-Graph Mapping (CSV / XLSX / SQLite)

Convert structured business tables directly into grounded RDF knowledge graphs matching an explicit RDFS/OWL ontology (TBox).

from safe_store import SafeStore, TBoxManager, TabularMapper

store = SafeStore(db_path="supply_chain.db")

# 1. Load TBox Ontology (Turtle format)
tbox = TBoxManager()
tbox.load_ontology("""
@prefix owl: <http://www.w3.org/2002/07/owl#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix ex: <http://example.org/ontology/> .

ex:Product a owl:Class .
ex:Supplier a owl:Class .

ex:suppliedBy a owl:ObjectProperty ;
    rdfs:domain ex:Product ;
    rdfs:range ex:Supplier .

ex:hasPrice a owl:DatatypeProperty ;
    rdfs:domain ex:Product .
""", format="turtle")

# 2. Declarative Mapping Configuration
mapping_rules = {
    "entity_mappings": [
        {
            "class": "http://example.org/ontology/Product",
            "subject_template": "http://example.org/product/{sku}",
            "properties": {
                "product_name": "http://example.org/ontology/hasName",
                "unit_price": "http://example.org/ontology/hasPrice"
            }
        },
        {
            "class": "http://example.org/ontology/Supplier",
            "subject_template": "http://example.org/supplier/{supplier_id}",
            "properties": {
                "supplier_name": "http://example.org/ontology/hasName"
            }
        }
    ],
    "relationship_mappings": [
        {
            "predicate": "http://example.org/ontology/suppliedBy",
            "source_template": "http://example.org/product/{sku}",
            "target_template": "http://example.org/supplier/{supplier_id}"
        }
    ]
}

# 3. Ingest CSV or Excel Sheet directly into ABox Graph
mapper = TabularMapper(store=store, tbox=tbox)
summary = mapper.map_csv("inventory.csv", mapping_rules=mapping_rules)
# Alternatively: mapper.map_excel("inventory.xlsx", mapping_rules=mapping_rules, sheet_name="Q3_Stock")
# Alternatively: mapper.map_sqlite_table("legacy.db", "products", mapping_rules=mapping_rules)

print(f"Mapped {summary['records_processed']} records into {summary['triples_generated']} RDF triples.")

4. Declarative Tabular-to-Graph Mapping (CSV / XLSX / SQLite)

Instant, Zero-LLM Knowledge Graph Construction from Structured Data

When your data already lives in structured tables (CSV exports, Excel sheets, or legacy SQLite databases), you don't need slow LLM-based extraction. safe_store provides a declarative mapping engine that transforms tabular records into grounded RDF knowledge graphs in milliseconds—no tokens consumed, no API latency.

How It Works

Define Your Ontology (TBox): Load an RDFS/OWL schema that describes your domain classes and relationships.
Write Mapping Rules: Declare how table columns map to entity classes, properties, and relationships using simple templates.
Bulk Ingest: Point the mapper at your file or database table. It performs batch transactional insertion directly into the graph store.

Complete Working Example

from safe_store import SafeStore, TBoxManager, TabularMapper

# 1. Initialize the store (no LLM required for mapping!)
store = SafeStore(db_path="supply_chain.db")

# 2. Load your domain ontology (Turtle format)
tbox = TBoxManager()
tbox.load_ontology("""
@prefix owl: <http://www.w3.org/2002/07/owl#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix ex: <http://example.org/ontology/> .

ex:Product a owl:Class .
ex:Supplier a owl:Class .

ex:suppliedBy a owl:ObjectProperty ;
    rdfs:domain ex:Product ;
    rdfs:range ex:Supplier .

ex:hasPrice a owl:DatatypeProperty ;
    rdfs:domain ex:Product .
ex:hasName a owl:DatatypeProperty .
""", format="turtle")

# 3. Define declarative mapping rules
mapping_rules = {
    "entity_mappings": [
        {
            "class": "http://example.org/ontology/Product",
            "subject_template": "http://example.org/product/{sku}",
            "properties": {
                "product_name": "http://example.org/ontology/hasName",
                "unit_price": "http://example.org/ontology/hasPrice"
            }
        },
        {
            "class": "http://example.org/ontology/Supplier",
            "subject_template": "http://example.org/supplier/{supplier_id}",
            "properties": {
                "supplier_name": "http://example.org/ontology/hasName"
            }
        }
    ],
    "relationship_mappings": [
        {
            "predicate": "http://example.org/ontology/suppliedBy",
            "source_template": "http://example.org/product/{sku}",
            "target_template": "http://example.org/supplier/{supplier_id}"
        }
    ]
}

# 4. Map your structured data instantly
mapper = TabularMapper(store=store, tbox=tbox)

# From CSV
summary = mapper.map_csv("inventory.csv", mapping_rules=mapping_rules)
print(f"CSV: {summary['records_processed']} rows -> {summary['entities_created']} entities")

# From Excel (specific sheet)
summary = mapper.map_excel("inventory.xlsx", mapping_rules=mapping_rules, sheet_name="Q3_Stock")

# From SQLite table
summary = mapper.map_sqlite_table("legacy.db", "products", mapping_rules=mapping_rules)

Input Format Examples

CSV Input (inventory.csv):

sku,product_name,unit_price,supplier_id,supplier_name
WIDGET-001,Industrial Widget,49.99,SUP-ACME,Acme Manufacturing
GADGET-002,Smart Gadget,199.99,SUP-TECH,TechParts Ltd

Mapping Templates:

{sku} → Replaced with value from the sku column
{supplier_id} → Replaced with value from the supplier_id column
Creates URIs like http://example.org/product/WIDGET-001

Why Use Tabular Mapping?

Feature	LLM Extraction	Tabular Mapping
Speed	Slow (seconds per chunk)	Instant (thousands of rows/sec)
Cost	High (API tokens)	Zero (local computation)
Accuracy	Probabilistic	Deterministic
Ontology Compliance	Prompt-dependent	Schema-enforced
Use Case	Unstructured text	Structured tables

Integration with SPARQL

Once mapped, your tabular data is immediately queryable via the full SPARQL 1.1 engine:

from safe_store import GraphStore

graph = GraphStore(store=store)

# Find all products supplied by Acme
results = graph.query_sparql("""
PREFIX ex: <http://example.org/ontology/>
SELECT ?productName ?price WHERE {
    ?product a ex:Product ;
             ex:hasName ?productName ;
             ex:hasPrice ?price ;
             ex:suppliedBy ?supplier .
    ?supplier ex:hasName "Acme Manufacturing" .
}
""")

for binding in results["results"]["bindings"]:
    print(f"Product: {binding['productName']['value']}, Price: {binding['price']['value']}")

5. Tri-Modal Unified Graph Retrieval (`query_graph_hybrid`)

Execute multi-channel queries combining Graph Subgraph Exploration, Dense Vectors, and Sparse BM25 Lexical search in a single call.

from safe_store import SafeStore, GraphStore

store = SafeStore(db_path="enterprise_kb.db", vectorizer_name="st")
graph = GraphStore(store=store)

# Unified retrieval: discovers related subgraph entities + BM25 hits + semantic vector chunks
response = graph.query_graph_hybrid(
    query_text="What microservices depend on AuthEngine and what database tables do they use?",
    top_k=5,
    dense_weight=0.4,
    bm25_weight=0.3,
    graph_weight=0.3
)

print(f"Retrieved {len(response['ranked_chunks'])} fused context chunks.")
print(f"Identified Subgraph Nodes: {len(response['subgraph']['nodes'])}")
print(f"Identified Subgraph Edges: {len(response['subgraph']['relationships'])}")

🔄 Database Portability & Re-Vectorization

safe_store allows you to migrate your entire knowledge base between different embedding models and export/import your database for backup or transfer.

Re-Vectorizing a Database

If you want to switch from one vectorizer (e.g., Sentence-Transformers) to another (e.g., OpenAI or Ollama), you can re-vectorize the entire database in-place. This will decrypt chunks, re-embed them using the new model, and update the database metadata atomically.

import safe_store

store = safe_store.SafeStore("my_knowledge.db", vectorizer_name="st")

# Re-vectorize using OpenAI's text-embedding-3-small
store.revectorize_database(
    new_vectorizer_name="openai",
    new_vectorizer_config={"model": "text-embedding-3-small"}
)

print("Database successfully migrated to OpenAI embeddings.")

Exporting and Importing Databases

You can export the entire state of your database (documents, chunks, vectors, graphs, FTS indices) to a portable JSON file. This is useful for backups, sharing datasets, or migrating between machines.

import safe_store

# 1. Export the database
store = safe_store.SafeStore("my_knowledge.db", vectorizer_name="st")
store.export_database("backup.json", decrypt=False) # Set decrypt=True to export plaintext

# 2. Import the database on another machine or into a new file
new_store = safe_store.SafeStore.import_database(
    input_path="backup.json",
    db_path="restored_knowledge.db",
    vectorizer_name="st" # Must match the exported vectorizer or be re-vectorized after import
)

Handling Encrypted Databases

If the database is encrypted, you can export it securely (keeping the encrypted blobs) or decrypt it during export. When importing, you must provide the decryption_key if the data was exported in its encrypted state.

# Export encrypted data (requires key to read, but keeps it encrypted in JSON)
store = safe_store.SafeStore("secure.db", encryption_key="secret123")
store.export_database("secure_backup.json", decrypt=False)

# Import encrypted data into a new encrypted store
new_store = safe_store.SafeStore.import_database(
    input_path="secure_backup.json",
    db_path="restored_secure.db",
    decryption_key="secret123", # Required to read the encrypted JSON blobs
    encryption_key="newsecret456" # Optional: re-encrypt with a new key
)

🔐 Zero-Leakage Encryption at Rest

safe_store provides transparent, chunk-level authenticated encryption using Fernet (AES-128-CBC with HMAC-SHA256). User-supplied passwords are hardened via PBKDF2-HMAC-SHA256 (600,000 iterations) before key derivation.

What Is Protected

Data	Encrypted?	Notes
Chunk text	✅ Yes	Decrypted transparently during `query()`
Document metadata	✅ Yes	JSON blob is encrypted at rest
Document full_text	✅ Yes	Stored in `documents` table
Vector embeddings	❌ No	Required for similarity search
Graph nodes/edges	❌ No	Structural knowledge graph data
File paths / timestamps	❌ No	Operational metadata

Basic Usage

import safe_store

# 1. Create an encrypted store
store = safe_store.SafeStore(
    db_path="classified.db",
    encryption_key="my-super-secure-passphrase",
    vectorizer_name="st",
    vectorizer_config={"model": "all-MiniLM-L6-v2"}
)

with store:
    # Document and metadata are encrypted before hitting SQLite
    store.add_document("confidential_contract.pdf", metadata={"classification": "Top Secret"})
    
    # Query decrypts chunks transparently in memory
    results = store.query("liability clauses", top_k=2)
    print(results[0]["chunk_text"])

Opening Without a Key (Graceful Degradation)

If the database is opened without providing the encryption key, queries still function but return encrypted placeholders instead of plaintext. This prevents accidental crashes while signalling that the data is protected.

# Re-open the same database WITHOUT the key
unauth_store = safe_store.SafeStore("classified.db", encryption_key=None)

with unauth_store:
    res = unauth_store.query("liability clauses", top_k=1)
    print(res[0]["chunk_text"])
    # >>> "[Encrypted Chunk - Key Unavailable]"

Wrong Key Detection

Supplying an incorrect key is detected immediately during decryption (via Fernet's HMAC verification). The library distinguishes between "no key provided" and "wrong key provided":

# Re-open with an INCORRECT key
wrong_store = safe_store.SafeStore(
    "classified.db",
    encryption_key="this-is-definitely-wrong"
)

with wrong_store:
    res = wrong_store.query("liability clauses", top_k=1)
    print(res[0]["chunk_text"])
    # >>> "[Encrypted Chunk - Decryption Failed]"

Verifying Encryption Programmatically

You can inspect the database directly to confirm that encryption flags are set correctly on every chunk and document:

import sqlite3

store = safe_store.SafeStore(
    "audit.db",
    encryption_key="audit-key",
    vectorizer_name="st"
)

with store:
    store.add_text("sensitive_unique_42", "Payload data here.", metadata={"owner": "Alice"})

# Verify raw DB state
conn = sqlite3.connect("audit.db")
cursor = conn.cursor()
cursor.execute("SELECT is_encrypted FROM chunks WHERE doc_id = 1")
flags = cursor.fetchall()
assert all(flag[0] == 1 for flag in flags), "Not all chunks are encrypted!"
conn.close()

Metadata Encryption

When encryption is enabled, the metadata dictionary is also encrypted as a single JSON blob. This is transparent during queries:

with store:
    store.add_text(
        unique_id="report_001",
        text="Q3 Financial Analysis...",
        metadata={"department": "Finance", "clearance": "Restricted"}
    )
    
    # The metadata is decrypted and prepended as context in query results
    results = store.query("Q3 analysis", top_k=1)
    print(results[0]["document_metadata"])
    # >>> {'department': 'Finance', 'clearance': 'Restricted'}

Security Considerations

Fixed Salt: This implementation uses a fixed salt for PBKDF2 derivation. This means the same password always yields the same key, which is a deliberate trade-off for portability (a single .db file can be moved between machines without external salt storage). For higher security requirements, consider wrapping the database file with OS-level full-disk encryption.
Vectors Remain Plaintext: Vector embeddings are stored as raw BLOBs to allow cosine-similarity search without decrypting the entire dataset. If your threat model requires vectors to be secret, encrypt the underlying filesystem.
Memory Safety: Decryption occurs in-memory during query(). Plaintext chunks exist only for the duration of the result formatting and are not cached outside of the SQLite connection scope.

Complete Example: Encrypted Document Lifecycle

import safe_store
from pathlib import Path
import shutil

DB_FILE = "encrypted_lifecycle.db"
KEY = "correct-horse-battery-staple"

# Cleanup from previous runs
for p in [DB_FILE, f"{DB_FILE}.lock", f"{DB_FILE}-wal", f"{DB_FILE}-shm"]:
    Path(p).unlink(missing_ok=True)

# Phase 1: Write encrypted data
writer = safe_store.SafeStore(
    db_path=DB_FILE,
    vectorizer_name="st",
    vectorizer_config={"model": "all-MiniLM-L6-v2"},
    encryption_key=KEY
)

doc = Path("secret_notes.txt")
doc.write_text("Project Phoenix launch is Q4. Key personnel: Alice, Bob.")

with writer:
    writer.add_document(doc, metadata={"sensitivity": "high"})
    print("Document encrypted and stored.")

# Phase 2: Read with correct key
reader = safe_store.SafeStore(DB_FILE, encryption_key=KEY)
with reader:
    results = reader.query("Project Phoenix", top_k=1)
    assert "Project Phoenix" in results[0]["chunk_text"]
    print("Decryption successful with correct key.")

# Phase 3: Read without key (placeholder)
no_key = safe_store.SafeStore(DB_FILE, encryption_key=None)
with no_key:
    res = no_key.query("Project Phoenix", top_k=1)
    assert res[0]["chunk_text"] == "[Encrypted Chunk - Key Unavailable]"
    print("Confirmed: no key returns placeholder.")

# Phase 4: Read with wrong key (tamper detection)
bad_key = safe_store.SafeStore(DB_FILE, encryption_key="wrong-key")
with bad_key:
    res = bad_key.query("Project Phoenix", top_k=1)
    assert res[0]["chunk_text"] == "[Encrypted Chunk - Decryption Failed]"
    print("Confirmed: wrong key is rejected via HMAC.")

# Cleanup
doc.unlink(missing_ok=True)
for p in [DB_FILE, f"{DB_FILE}.lock", f"{DB_FILE}-wal", f"{DB_FILE}-shm"]:
    Path(p).unlink(missing_ok=True)
print("Encrypted lifecycle demo complete.")

📊 Semantic Datalake & Point Cloud Visualization

safe_store includes a powerful Semantic Datalake Engine that allows you to visualize your entire knowledge base as an interactive 2D or 3D point cloud. This is essential for understanding data clustering, identifying outliers, and auditing the quality of your embeddings.

1. Programmatic Projections (PCA, t-SNE, UMAP)

You can reduce the high-dimensional vectors to 2D or 3D coordinates using PCA (Principal Component Analysis) or t-SNE (t-Distributed Stochastic Neighbor Embedding).

import safe_store

store = safe_store.SafeStore("my_knowledge.db", vectorizer_name="st")

with store:
    # Get 2D PCA projection as a list of dictionaries
    points_2d = store.get_datalake_view(
        method='pca', 
        n_components=2, 
        output_format='dict'
    )

    for p in points_2d:
        print(f"Doc: {p['document_title']} | X: {p['x']:.2f}, Y: {p['y']:.2f}")

    # Get 3D t-SNE projection
    points_3d = store.get_datalake_view(
        method='tsne', 
        n_components=3, 
        output_format='dict'
    )

2. Persistent Caching for Instant Visualization

Projecting 100,000+ vectors can take several seconds. safe_store automatically caches the projection results inside the SQLite database. The next time you call get_datalake_view with the same parameters, it returns instantly.

# First call: Computes PCA and caches results (takes ~2s)
store.get_datalake_view(method='pca', use_cache=True)

# Second call: Returns instantly from SQLite cache
store.get_datalake_view(method='pca', use_cache=True)

Note: Cache is automatically invalidated whenever a document is added or deleted.

3. Lazy Streaming for Massive Datasets

For extremely large databases that might not fit into RAM, use the lazy streaming generator. It uses IncrementalPCA to process vectors in batches.

# Stream points in batches of 500
stream = store.stream_datalake_chunks(batch_size=500, method='incremental_pca')

for point in stream:
    # Process point-by-point without loading the whole matrix
    print(point['x'], point['y'])

4. Interactive HTML Visualizer Export

The most powerful feature is the ability to export a standalone, interactive HTML file that you can share with others. It includes a Plotly-powered 3D canvas, search filtering, and a chunk inspector.

store.export_datalake_html(
    output_file="my_datalake.html",
    title="Enterprise Knowledge Base Audit",
    method='tsne',
    n_components=3
)

Features of the exported HTML:

Interactive 3D/2D Canvas: Rotate, zoom, and pan through your data.
Hover Inspection: See the actual text content and metadata of any point.
Real-time Filtering: Filter points by document title or metadata keywords.
Zero Dependencies: The exported file works offline in any modern browser.

🎯 Supported Vectorization Backends

Backend	Identifier	Typical Model / Target	Local / Remote
Sentence-Transformers	`"st"`	`all-MiniLM-L6-v2`, `all-mpnet-base-v2`	Local (PyTorch / HuggingFace)
Ollama	`"ollama"`	`nomic-embed-text`, `qwen3-embedding`	Local (Ollama Server)
OpenAI	`"openai"`	`text-embedding-3-small`, `text-embedding-3-large`	Remote API
Cohere	`"cohere"`	`embed-english-v3.0`, `embed-multilingual-v3.0`	Remote API
Lollms	`"lollms"`	Any OpenAI-compatible local/remote endpoint	Local / Remote
TF-IDF	`"tfidf"` / `"tf_idf"`	Data-dependent sparse baseline	Local (Scikit-Learn)
Grepper	`"grepper"`	Lightweight inverted index with markdown trees	Local (Zero-ML)

📑 Supported Document & File Formats

safe_store parses structured, unstructured, and source files out-of-the-box:

Unstructured Documents: .pdf, .docx, .pptx, .html, .htm, .txt, .md, .rst, .msg, .rtf
Data & Tables: .csv, .tsv, .json, .xlsx, .xls, .xml, .sql
Source Code: .py, .js, .ts, .tsx, .jsx, .c, .cpp, .h, .cs, .java, .go, .rs, .php, .rb, .swift, .kt, .sh, .ps1, .lua, .sql

🔍 W3C SPARQL 1.1 Query Forms Cheat Sheet

safe_store natively executes all four standard W3C SPARQL 1.1 query forms across your knowledge graph:

Query Form	Purpose	Return Type	Typical Use Case
`SELECT`	Tabular projections across graph patterns	`{"head": {"vars": [...]}, "results": {"bindings": [...]}}`	Relational multi-hop traversals, aggregations (`COUNT`, `GROUP BY`), and filtered lookups.
`ASK`	Boolean existence test	`{"boolean": True / False}`	Fast sanity checking and compliance verification without retrieving payloads.
`CONSTRUCT`	Subgraph transformation & inference	`{"triples": [{"subject": ..., "predicate": ..., "object": ...}]}`	Transforming schemas, creating direct shortcut edges, or exporting custom RDF subgraphs.
`DESCRIBE`	Resource neighborhood extraction	`{"triples": [...]}`	Pulling all known incoming and outgoing triples associated with an entity.

📊 Performance Benchmarks

Typical benchmarks measured on consumer hardware (Intel i7 / 16GB RAM / SSD):

Operation	Scale / Dataset	Elapsed Time	Mode
Dense Vector Query	50,000 Chunks	~15 ms	NumPy Cosine Dot Product
BM25 Lexical Search	100,000 Chunks	~4 ms	SQLite FTS5 (Porter Stemmed)
W3C SPARQL Relational Join	20,000 Triples (2-hop)	~8 ms	RDFLib + In-Memory Quad Index
Tabular Mapping	10,000 CSV Rows	~1.2 s	Batch Transactional Insertion
Document Ingestion (ST)	1 MB Text (~300 pages)	~3.5 s	Parsing + Token Chunking + Embedding

6. The 8 RAG Chunking Strategies (Beyond the Basics)

Retrieval quality is decided at cut time. safe_store implements a complete suite of 8 distinct chunking strategies:

 1. Fixed-Size [====][====][====]  -> Slices at fixed intervals (fast, baseline)
 2. Overlap    [====--]            -> Rescues broken sentences across boundaries
                  [--====--]
 3. Recursive  Document            -> Splits paragraphs -> sentences -> words
               ├── Para 1
               └── Para 2 -> S1, S2
 4. Semantic   ───📉───📉───       -> Cuts at cosine similarity valleys (topic shifts)
 5. Contextual [Prefix] + [Chunk]  -> Prepends full-document situating context (Anthropic)
 6. Structure  # H1 > ## H2        -> Injects section breadcrumb paths [H1 > H2]
 7. Late       Tokens ──[Transformer]──> Contextual Embeddings ──[Mean Pool]──> Vectors
 8. Graph      Entities & Relations-> Tri-Tier Multi-Hop Graph Traversal

Strategy	Flag	Ideal For	Mechanics & Key Benefit
Token Window	`'token'` (Default)	Standard RAG	Slices by tokenizer limits (`tiktoken`/HF) with offset mapping preserving all `\n` line breaks.
Recursive Tree	`'recursive'`	General Docs & Code	Hierarchically splits by `\n\n` $\rightarrow$ `# Headers` $\rightarrow$ `\n` $\rightarrow$ sentences $\rightarrow$ words. Best all-around balance.
Structure-Aware	`'structure'` / `'markdown'`	Technical Manuals & Specs	Parses Markdown `# H1` $\rightarrow$ `## H2` $\rightarrow$ `### H3` stacks, attaching lineage breadcrumbs `[H1 > H2]`.
Semantic Valley	`'semantic'`	Long Essays & Narrative	Embeds sentences and cuts where adjacent cosine similarity drops below threshold (topic boundary).
Contextual Retrieval	`'contextual'`	Complex Knowledge Bases	Injects full-document situating summaries before storage (Anthropic pattern), eliminating ambiguous pronouns.
Late Chunking	`'late'`	Dense Technical Context	Passes the entire document through the transformer first, then mean-pools chunk token representations (Jina AI pattern).
Paragraph	`'paragraph'`	Articles & Prose	Groups double-newline paragraph blocks up to `chunk_size` without mid-thought cuts.
Fixed Character	`'character'`	Raw Log Streams	Fast character slicing with sliding window overlap.

Strategy Implementation Examples

from safe_store import SafeStore

# Strategy A: Structure-Aware Markdown with Breadcrumbs
store_md = SafeStore(
    "manual.db",
    vectorizer_name="st",
    chunk_size=200,
    chunking_strategy="structure" # Injects [Section: Architecture > Storage > WAL] into chunks
)

# Strategy B: Semantic Chunking (Topic Shift Detection)
store_sem = SafeStore(
    "research.db",
    vectorizer_name="st",
    chunk_size=300,
    chunking_strategy="semantic", # Splits at cosine similarity valleys
    chunking_kwargs={"similarity_threshold": 0.65}
)

# Strategy C: Contextual Retrieval (Anthropic Pattern)
def my_context_enricher(full_doc: str, chunk: str) -> str:
    # Optional LLM or heuristic summary
    return f"From document '{full_doc[:40]}...': Topic covers database storage engine."

store_ctx = SafeStore(
    "enterprise.db",
    vectorizer_name="st",
    chunk_size=256,
    chunking_strategy="contextual",
    context_enricher=my_context_enricher
)

# Strategy D: Context Expansion Windowing
store_exp = SafeStore(
    "logs.db",
    vectorizer_name="st",
    chunk_size=128,
    expand_before=30, # Injects 30 tokens of preceding context into LLM prompt
    expand_after=30   # Injects 30 tokens of succeeding context into LLM prompt
)

🗺️ Roadmap

🤝 Contributing & License

Contributions are welcome! Please open an issue or submit a pull request on GitHub.

Licensed under the Apache 2.0 License.

Name		Name	Last commit message	Last commit date
app/api/v1		app/api/v1
docs		docs
examples		examples
point_cloud_web_app		point_cloud_web_app
safe_store		safe_store
scripts		scripts
src/commands/chatPanel/webview		src/commands/chatPanel/webview
tests		tests
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
DOC.md		DOC.md
LICENSE		LICENSE
README.md		README.md
diagnose_tests.py		diagnose_tests.py
package.json		package.json
plan.md		plan.md
pyproject.toml		pyproject.toml
todo.md		todo.md

Latest commit

History

Folders and files

Repository files navigation