AI/LLM Configuration

The AI/LLM Configuration section contains the administrative settings used to configure, integrate, and manage AI and Large Language Model functionality in elDoc.

elDoc uses an LLM-agnostic architecture, allowing different providers and models to be configured for different AI tasks.

Depending on deployment requirements, elDoc can use cloud-based AI services, locally hosted models, OpenAI-compatible APIs, AWS Bedrock, and other supported integrations.

Administrators can configure different models for different functional roles rather than requiring a single model to perform every AI operation.

AI Providers

AI providers define the external or locally hosted services through which elDoc accesses AI models.

Depending on the deployment, providers can include:

  • OpenAI-compatible APIs;

  • AWS Bedrock;

  • locally deployed inference services;

  • supported cloud AI providers.

Provider configuration can include endpoint information, authentication credentials, regions, model identifiers, and other provider-specific settings.

Model Configuration

Administrators can configure the models available to elDoc and assign them to specific AI functions.

Different models can be used for:

  • chat;

  • reasoning;

  • AI Agents;

  • tool usage;

  • AI Document Processing;

  • OCR and vision processing;

  • embeddings;

  • reranking.

This allows the AI architecture to be optimized for accuracy, performance, latency, infrastructure requirements, and cost.

Chat Models

Chat models are used for natural-language interaction through GenAI Chat.

The configured model processes user requests and generates responses using the context provided by elDoc retrieval and RAG functionality.

A dedicated model can be selected for Chat independently from models used for other AI tasks.

Reasoning Models

Reasoning models can be configured for AI operations that require more complex planning, analysis, or multi-step reasoning.

These models can be used by Agentic RAG and AI Agents when a request requires more than a direct LLM response.

Tool-Usage Models

elDoc supports models capable of tool usage.

Tool-capable models can invoke predefined elDoc functions to perform actions such as retrieving information, working with files, or executing document-management operations.

A dedicated tool-use model can be configured separately from the primary reasoning model.

AI Document Processing Models

AI Document Processing can use dedicated LLM or multimodal models for:

  • document classification;

  • structured data extraction;

  • form processing;

  • document understanding;

  • OCR;

  • vision-based processing.

Different models can be selected according to the document-processing requirements of the organization.

OCR and Vision Models

AI/LLM-based OCR and image understanding can be configured using models capable of processing visual document content.

These models can be used to extract text or structured information from:

  • scanned documents;

  • image-based PDFs;

  • photographs;

  • other image-based files.

Extracted information can then be indexed and used by other elDoc AI capabilities.

Embedding Models

Embedding models generate vector representations of document content.

These vectors are used for:

  • semantic search;

  • similarity retrieval;

  • RAG;

  • Agentic RAG.

Embedding models can be provided by supported cloud services or locally deployed through compatible APIs.

Vector Indexing

Vector indexing configuration determines how document embeddings are generated and stored for semantic retrieval.

Vector indexing can be applied to digitally generated documents as well as image-based files after their content has been extracted.

The vector index forms the semantic retrieval layer used by GenAI Chat, RAG, and Agentic RAG.

Keyword and Hybrid Search

elDoc can combine vector retrieval with traditional keyword search.

Apache Solr is used as part of the retrieval architecture for keyword and full-text search and, depending on the configured deployment, additional retrieval scenarios.

Hybrid retrieval combines:

  • full-text relevance;

  • exact keyword matches;

  • vector similarity;

  • reranking.

This provides a stronger retrieval foundation for RAG-based AI processing.

Reranking Models

Reranker models evaluate the relevance of retrieved documents or document fragments.

The reranking stage is performed after initial retrieval and before the selected information is provided to the primary LLM.

elDoc supports dedicated reranking/scoring models through supported integrations, including AWS Bedrock and Cohere-compatible capabilities.

AI Indexing Configuration

AI-powered Document Indexing uses configured AI models to extract structured values from documents.

AI extraction uses information defined in:

  • Indexing Profiles;

  • Indexing Fields;

  • field data types;

  • field-level AI extraction context;

  • profile-level AI extraction context.

Administrators configure Document Indexing separately under:

AI File Management Configuration → Document Indexing Configuration

The AI/LLM configuration determines which models are available to perform the corresponding extraction tasks.

Local Model Integration

elDoc can integrate with locally deployed models through OpenAI-compatible APIs.

Local model integration can be used for:

  • chat;

  • reasoning;

  • tool usage;

  • embeddings;

  • reranking;

  • AI Document Processing;

  • other supported AI functions.

This enables private deployments in which document content remains within the organization's own infrastructure.

AWS Bedrock Integration

AWS Bedrock can be used as an AI provider for multiple elDoc capabilities.

Depending on the configured models and Bedrock services, it can provide:

  • LLM inference;

  • embeddings;

  • reranking/scoring;

  • document-processing models;

  • other supported AI functions.

Administrators can select Bedrock models independently for different AI tasks.

Model Assignment by Function

elDoc allows models to be selected according to their intended functional role.

For example:

Chat                     → Model A
Reasoning                → Model B
AI Agents                → Model B
Tool Usage               → Model C
Embeddings               → Model D
Reranking                → Model E
AI Document Processing   → Model F
OCR / Vision             → Model G

The same model can be assigned to multiple functions, or different specialized models can be used for each task.

This architecture allows organizations to balance:

  • response quality;

  • inference speed;

  • GPU or CPU requirements;

  • cost;

  • token consumption;

  • security;

  • model capabilities.

AI Usage and Billing

elDoc supports accounting and billing for AI-related usage.

AI consumption can be tracked for supported operations, including AI processing and indexing.

This provides a common mechanism for monitoring and controlling AI usage across the platform.

Security and Access Control

AI/LLM configuration does not bypass the elDoc security model.

Document access during:

  • GenAI Chat;

  • semantic retrieval;

  • RAG;

  • Agentic RAG;

  • AI Agents;

  • AI Indexing;

remains subject to the current user's permissions.

Only documents accessible to the corresponding user can be included in retrieval and AI processing.

Deployment Models

The AI architecture supports different deployment approaches, including:

  • cloud-only;

  • local/on-premises;

  • private cloud;

  • isolated or air-gapped environments;

  • hybrid deployments combining local and cloud models.

Administrators can select the architecture that best matches organizational requirements for data residency, confidentiality, performance, availability, and cost.

Last modified: August 26, 2026