Vectorless

Reasoning-based document indexing without vector databases

Vectorless extracts hierarchical document structure from PDFs and other documents using AI agents, enabling intelligent document navigation and retrieval without vector databases or chunking.

Features

• No Vector DB - Uses document structure and LLM reasoning instead of vector similarity
• No Chunking - Documents organized into natural sections, not artificial chunks
• Human-like Retrieval - Simulates how experts navigate complex documents
• Multi-Document Search - Search across multiple documents with cross-document relations
• Efficient Tree Search - Hybrid MCTS/Greedy algorithm for optimal performance
• Expert Knowledge - Inject domain expertise without fine-tuning
• Multi-Node Extraction - Extract and aggregate content from multiple nodes
• Streaming SSE - Real-time progress events during processing
• Provider Agnostic - Works with any AI SDK v6 compatible model

Installation

npm install vectorless
# or
pnpm add vectorless
# or
yarn add vectorless

Quick Start

import { createTreeSearchOrchestrator, getDomainTemplate } from "vectorless";
import { openai } from "@ai-sdk/openai";

const model = openai("gpt-4o-mini");
const orchestrator = createTreeSearchOrchestrator(model);

// Search with automatic algorithm selection
for await (const { event, result } of orchestrator.search(
  "What are the key findings?",
  knowledgeBase.tree,
)) {
  console.log(event.type, event.data);
  if (result) {
    console.log("Answer:", result.extractedContent);
  }
}

Multi-Document Search

import { createMultiDocSearchAgent } from "vectorless";

const multiSearch = createMultiDocSearchAgent(model);

for await (const { result } of multiSearch.searchAcross(
  "Compare conclusions across documents",
  [kb1, kb2, kb3],
  { aggregationStrategy: "merge" },
)) {
  if (result) {
    console.log("Merged answer:", result.mergedContent);
    console.log("Cross-doc relations:", result.crossDocRelations);
  }
}

Domain Templates

import { getDomainTemplate, createMemoryPreferenceStore } from "vectorless";

const store = createMemoryPreferenceStore();
const legalTemplate = await store.getTemplate("legal");
// Templates: legal, medical, technical, academic, financial

Architecture

Vectorless follows hexagonal architecture (ports & adapters pattern):

src/
├── domain/           # Pure business logic (schemas, types)
├── ports/            # Interface definitions (contracts)
├── agents/           # AI agents (20+ specialized agents)
│   ├── query-classifier.ts    # Query complexity detection
│   ├── greedy-search.ts       # Fast O(log n) search
│   ├── mcts-search.ts         # Monte Carlo Tree Search
│   ├── multi-doc-search.ts    # Cross-document search
│   ├── content-extractor.ts   # Fragment extraction
│   └── multi-node-aggregator.ts # Content aggregation
├── use-cases/        # Application logic orchestrators
├── templates/        # Domain-specific templates
└── infrastructure/   # Default adapters (memory, file)

Core Agents

Agent	Purpose
QueryClassifierAgent	Classifies query as simple/complex/multi-doc
GreedySearchAgent	BM25-based O(log n) tree traversal
MCTSSearchAgent	UCB1-based Monte Carlo Tree Search
MultiDocSearchAgent	Parallel search across knowledge bases
ContentExtractorAgent	Extracts relevant fragments from nodes
MultiNodeAggregatorAgent	Deduplicates and merges content
TocDetectorAgent	Detects document table of contents
StructureExtractorAgent	Builds hierarchical tree from TOC
SummarizerAgent	Generates section summaries
EntityExtractorAgent	Extracts named entities
RelationExtractorAgent	Finds entity relationships
QuestionAnswerAgent	Answers questions from KB

Tree Search Algorithms

Query Classification

const classifier = createQueryClassifierAgent(model);
const classification = await classifier.classify("What year was founded?", 1);
// { type: "simple", confidence: 0.9, suggestedAlgorithm: "greedy" }

Greedy Search (Simple Queries)

• Complexity: O(log n)
• Scoring: BM25/TF-IDF + LLM refinement
• Best for: Single-fact lookups, direct answers

MCTS Search (Complex Queries)

• Algorithm: UCB1 with beam width
• Formula: score = Q/N + C * sqrt(ln(T)/N)
• Best for: Multi-step reasoning, comparisons

Multi-Document MCTS

• Parallel expansion across documents
• Aggregation strategies: merge, rank, cluster
• Cross-document relation detection

Domain Templates

Built-in templates for specialized domains:

Template	Focus
legal	Contracts, regulations, liability clauses
medical	Clinical trials, diagnoses, drug interactions
technical	API docs, specifications, code examples
academic	Research papers, citations, methodology
financial	Reports, metrics, risk factors

import { LEGAL_TEMPLATE, FINANCIAL_TEMPLATE } from "vectorless";

console.log(LEGAL_TEMPLATE.priorityKeywords);
// ["shall", "hereby", "pursuant", "liability", ...]

console.log(FINANCIAL_TEMPLATE.sectionWeights);
// { "financial highlights": 1.5, "risk factors": 1.3, ... }

Ports (Interfaces)

Vectorless defines clean interfaces for all external dependencies:

Port	Purpose
PdfParserPort	PDF text extraction
CachePort	Caching layer
KnowledgeBaseRepositoryPort	KB persistence
GreedySearchPort	Greedy search algorithm
MCTSSearchPort	MCTS search algorithm
QueryClassifierPort	Query classification
ContentExtractorPort	Content extraction
MultiNodeAggregatorPort	Content aggregation
PreferenceStorePort	User preferences & templates

Integration with OneGenUI

For UI components and full integration:

pnpm add @onegenui/vectorless

import { useTreeSearch, useMultiDocSearch } from "@onegenui/vectorless";

function SearchPanel() {
  const { search, results, isSearching } = useTreeSearch(model);
  // ...
}

License

Credits