GENAPA Hub Generally applicable platform assistance
Features

Everything GENAPA does.

Browse by area of concern, or jump to a section from the sidebar.

Search

Find anything by describing what you need.

Three ways to search your archive -- all driven by meaning, not keywords.

Seeker (multi-pass agentic RAG) The headline search capability. An AI search agent that decomposes your question into multiple focused searches, runs them iteratively (up to 5 passes), evaluates intermediate results, and refines until it has thorough coverage. This is not a single vector lookup -- it is an autonomous search agent that understands what you are looking for and keeps digging until it finds it.
Semantic search Fast vector-based search across file names, descriptions, claims, and content simultaneously. Two scoring strategies: best single field match or balanced scoring across all fields. Filter by node type, domain, collection, or job.
Natural conversation Chat with your archive using the Omni agent. No query syntax, no context setup -- just describe what you need. Answers are always grounded in your actual files, with citations you can follow back to the source.

Technical detail

Seeker uses a manual tool-call loop with SearchMemory chaining across iterations. Vector search powered by pgvector with HNSW indexes. Cross-field scoring with configurable per-field weighting. Results are relevancy-ranked by the LLM, not just by vector distance.

Why it matters:

You talk naturally and GENAPA finds what you need across everything -- even when you do not remember the right terms or where you saved something.

Processing

Optional layers of understanding. You choose the depth.

Every processing layer is independently configurable. Start with the baseline and add depth when your use case calls for it.

01
Baseline: Chunking + Source Field Extraction Files are split into searchable chunks with vector embeddings. AI generates summaries, descriptions, and key facts for each file. This alone gives you powerful search and AI chat across your entire archive. No complex setup required.
02
Deeper understanding: Synth Extraction + Synthesis Enable the Weaver agent to read across your sources and build a structured conceptual map -- like an auto-generated table of contents for your knowledge. The Oracle agent then generates detailed narrative descriptions for each concept. This gives you higher-order understanding of how your content relates.
03
Targeted enrichment: Reference + Entity Extraction For specialized use cases, enable extraction of specific named entities -- people, systems, concepts -- with cross-source deduplication. Useful when you need to track specific things across many documents. More resource-intensive, so enable it when the use case calls for it.

Technical detail

Version-based change tracking per operation ensures only modified content triggers re-processing. Domain-specific extraction prompts defined in YAML. Two-stage entity disambiguation: vector similarity candidates, then LLM evaluation with domain-specific rules.

Why it matters:

Start with fast, powerful search. Add depth when you need it. Every layer is independently configurable, and you control which AI model handles each task.

Agents

Purpose-built AI agents, each with a specific role.

Four specialized agents that handle different kinds of knowledge work -- and compose together.

Seeker Multi-pass search agent. Breaks your question into focused searches, runs them iteratively, evaluates results, and refines until coverage is thorough. The engine behind GENAPA's natural language search.
Oracle Deep exploration and synthesis agent. Discovers relevant content via Seeker, then produces written analysis with citations grounded in your source files.
Weaver Conceptual organization agent. Reads across your sources, identifies themes and groupings, and builds a structured hierarchy. Performs differential updates when content changes -- only affected areas are re-processed.
Omni Interactive chat agent. Understands your archive, loads skills dynamically, and delegates to other agents for deeper research. Supports sub-agent spawning and human-in-the-loop input.

Beyond the core agents, GENAPA includes a composable skills system with 19+ skills that can be loaded dynamically. Custom agent identities and personas let you tailor how agents interact with your content.

Technical detail

Tool-calling loop pattern with persistent session memory. Cross-agent memory linking preserves the chain of discovery. Sub-agent spawning with configurable recursion depth. Skill discovery and composition via semantic search.

Why it matters:

These are specialized knowledge workers, not generic chatbots. Each handles a specific kind of knowledge work, and they compose -- Omni can use any other agent as a tool.

Privacy

Your knowledge archive stays on your machine. Always.

Data sovereignty is built into the architecture, not bolted on as an afterthought.

  • Your entire knowledge archive — every extracted entity, relationship, synthesis, conversation, and memory — lives in a local database on your machine. GENAPA the company never has access to it.
  • The database stores metadata and descriptions -- not your actual file content. Content is retrieved on-demand from your filesystem when needed.
  • Choose from 10 AI providers and connect with your own API key — GENAPA never proxies through its servers. For fully air-gapped operation, run Ollama locally with zero external network calls.
  • All outputs stay on your machine and belong to you. Extracted knowledge, synthesized insights, embeddings, and search indexes are generated and stored locally. Nothing is sent back to GENAPA.
  • Docker security hardening: read-only containers, capability restrictions, no-new-privileges.
  • Optional usage telemetry (aggregate source counts, token usage, and version) can be sent to genapa.com for license management. No file content, knowledge, or archive data is ever included. Disable it entirely with a single configuration setting.
  • Encrypted credential storage (DPAPI on Windows). API key authentication with scoped access.
  • Passphrase-encrypted configuration export and import for secure portability.

Technical detail

Docker containers run with read-only filesystem, cap_drop ALL, no-new-privileges, and tmpfs for temp directories. DPAPI-encrypted LINK credentials. Fixed-time secret comparison prevents timing attacks. Scoped API keys (Admin, Mcp, Portal) restrict access per endpoint group.

Why it matters:

Most AI tools require your data on their servers to function. GENAPA keeps your knowledge archive entirely local — only the AI provider you choose sees the content you send for processing, and you can go fully offline at any time. During extraction and chat, relevant content from your files is sent directly to your provider's API — it never passes through GENAPA's infrastructure. Regulatory alignment (GDPR, EU AI Act) is built into the architecture: no user content is stored or processed on GENAPA's infrastructure.

Performance

Built for speed and scale.

Process thousands of files without re-analyzing your entire archive every time something changes.

  • Incremental processing: Version-based change tracking means only modified content is re-processed. Edit a file and only the changed parts are re-analyzed -- the rest of your archive is untouched.
  • Lightweight footprint: The database stores metadata and descriptions, not file content. Content is retrieved on-demand via GENAPA Link. This keeps the database compact enough to support large-scale archives.
  • Smart accumulation: File watchers batch changes with configurable debounce timers that reset when you are actively editing, preventing processing during active work sessions.
  • Write coalescing: Rapid memory updates are merged to prevent filesystem thrashing.
  • Connection pooling: Configurable PostgreSQL connection pooling with lifetime management.
  • Distributed locking: Redis-based locking with automatic renewal for safe concurrent processing.

Technical detail

Per-operation version counters on source nodes. Content-addressed chunks via SHA256 for change detection. Accumulation timers with reset-on-edit. Redis distributed locks with configurable expiration and renewal intervals.

Why it matters:

Process thousands of files without re-analyzing your entire archive every time something changes. GENAPA is designed to be efficient with both your time and your AI costs.

Configuration

Every aspect is configurable. Any knowledge domain.

No vendor lock-in. No one-size-fits-all processing. You control exactly how your content is analyzed.

  • 10 LLM providers: Azure OpenAI, OpenAI, Ollama (local), Anthropic, Google Vertex/Gemini, Amazon Bedrock, Hugging Face, Mistral, DeepSeek, xAI (Grok).
  • Per-feature model assignment: Use a powerful model for synthesis and a fast, cheap model for chunking. Each processing feature can use a different provider and model.
  • Domain-configurable extraction: Built-in domains for source code, technical documentation, and general content. Add new domains by writing a YAML configuration -- no code changes required.
  • Any knowledge domain: Source code, research papers, legal documents, personal notes, creative projects -- customize extraction instructions to pull out what matters for your domain.
  • Per-job configuration: Each processing job can have different settings, operations, and domain types.
  • Spend controls: Usage tracking and configurable spend caps with warning thresholds.

Technical detail

LlmFeature enum maps each feature to an independent provider/model configuration. Domain types carry extraction guidance, disambiguation rules, worked examples, field specifications, and synth type declarations. All defined in YAML.

Why it matters:

No vendor lock-in. No one-size-fits-all processing. You control exactly how your content is analyzed and which AI models do the work.

Integration

Connects to your existing workflow.

Your archive fits into your tools -- editors, scripts, automation -- not the other way around.

MCP (Model Context Protocol) Your archive becomes a tool that AI assistants in your editor can use. Search, retrieve, explore, and run agents from Claude Code, GitHub Copilot, or any MCP-compatible client.
REST API Full programmatic access -- 20+ endpoint groups covering chat, search, jobs, memory, configuration, and more.
GENAPA Link Native filesystem bridge that monitors your files for changes. Real-time detection with smart event coalescing -- rapid saves are batched, rename chains are simplified, and mid-edit processing is prevented.
File watchers Glob-based include/exclude patterns. Health monitoring. Configurable accumulation windows. Automatic orphan detection for deleted files with grace periods.
SignalR Real-time streaming for chat responses, archive updates, and job status notifications.
Docker deployment Install Docker, run the deployment script. That is it.

Technical detail

MCP over Streamable HTTP at /mcp. SignalR hubs for chat streaming and archive updates. LINK uses FileSystemWatcher with System.Reactive coalescing. Redis-backed job queuing with distributed locking.

Why it matters:

GENAPA fits into your existing tools -- editors, scripts, CI/CD. You do not have to switch to the portal to use your knowledge archive.

CLI

The full GENAPA platform, available as a shell command.

The GENAPA CLI (genapa) exposes everything the portal does -- plus things it doesn't. Submit ingestion jobs from a JSON file. Run the Seeker agent from a bash script. Retrieve any memory node by key from any machine. Cross-platform native binaries ship in the release bundle. No separate install. No package manager. If you prefer the portal, see the full web interface →

Automation and scripting Submit knowledge ingestion jobs from JSON files, poll job status, and trigger processing from cron jobs, build systems, or any shell-scriptable workflow. Every command supports --json for machine-readable, pipeline-composable output.
AI agents from the terminal genapa seek <query> runs the full multi-pass SEEKER discovery agent. genapa explore <question> runs ORACLE for written synthesis. Full agent intelligence, not a simplified API wrapper.
Secure, deployment-bound credentials genapa auth pair uses a challenge/response pairing flow that binds credentials to your deployment. Credentials are locally encrypted; all subsequent commands authenticate automatically. The --api-key option is also available for scripted or CI environments where interactive pairing isn't possible.
Memory access from anywhere genapa memory content <key> retrieves any memory node by GUID key from any authenticated machine -- search results, agent session records, or task context, in any environment with CLI credentials.
Multi-environment and headless-ready --environment dev|test|prod routes to the right deployment with a single flag. In server-only or containerized environments where no browser is available, the CLI is the full operational surface.

Technical detail

Commands: auth pair, seek, explore, search, memory content, memory relations, collections list, jobs list/get/create/cancel/stop/resume. All data commands support --json for machine-readable output. Global options: --environment, --api-url, --api-key. Auth pairing supports single-step, split begin/complete, and resume workflows. Native binaries per platform -- Windows, Linux, macOS x64, macOS ARM -- ship in the release bundle without requiring .NET SDK on the target machine.

Why it matters:

Everything the portal does through the browser is available as a composable shell command. For teams running GENAPA in CI/CD, server environments, or automation workflows, the CLI is not a convenience -- it is the primary interface. Memory nodes are accessible from the CLI -- retrieve any saved context by key with genapa memory content <key>. See the Memory section.

Download See all integration options
Memory

Context that travels. Memory that's searchable.

GENAPA memory is not preference tracking -- it is structured context you can save at runtime, address by key, and retrieve from any session, any machine, any agent role. Every memory enters the same embedding and indexing pipeline as your documents, so Seeker surfaces relevant memories during discovery alongside actual sources. No other major AI platform makes its memory searchable in the same knowledge base. Pass a single key and unlock any amount of typed, formatted context.

Save anything at runtime. Retrieve it anywhere. Call SaveMemory with a description and any JSON or text payload. You get back a GUID key. Retrieve the full content from any authenticated session, in any subsequent conversation, or from the CLI with genapa memory content <key>. No separate database. No external storage.
Key-based inter-agent handoff Orchestration without shared state: a planner agent saves a task package and returns its key. A worker agent receives the key and retrieves the full plan. A reviewer gets both. No live thread, no context window reuse -- keys are the coupling, and they work across time, machines, and agent roles.
Memories are part of your knowledge archive Every saved memory is written to a YAML file and immediately enters the chunking and embedding pipeline. When you run a Seeker query, relevant memories surface alongside documents -- not because you pointed the system there, but because the content is semantically indexed.
Typed hierarchy, not flat storage GenericMemory for ad hoc saves. SearchMemory for result sets with per-node relevancy. SeekerMemory, OracleMemory, and WeaverMemory for full agent session records with timeline, tool use, and cross-references. Each type formats and filters differently at retrieval.
Dual-TTL control and transparent storage Two independent expiration controls: cache TTL and persistence TTL. Set them independently, or omit for indefinite retention. Stored as YAML on your local disk -- readable, portable, and auditable. The startup rehydration service restores in-memory cache automatically after restarts.

Technical detail

Typed memory hierarchy: GenericMemory, SearchMemory, NodeMappingMemory, SeekerMemory, OracleMemory, WeaverMemory, OmniMemory. Every write goes to a YAML file via an atomic temp-file-then-rename operation with 500ms write coalescing. On retrieval, cache is checked first; misses load from disk and re-cache. Startup rehydration loads all non-expired files in parallel. MemoryToMemory graph relations are materialized in the knowledge store after each disk write -- memory chains are navigable as graph edges. Bulk save, retrieve, and delete in single calls.

Why it matters:

Most AI memory systems remember facts about you. GENAPA memory carries structured work -- task packages, search results, agent session records, synthesis outputs -- across any number of sessions, agents, and machines. The key-based handoff pattern is what makes multi-agent orchestration tractable: no shared state, no context window limits, no coordination infrastructure beyond a GUID string.

Explore the CLI See all features
Visualization

See the shape of your knowledge.

Understanding your content is not just about text results -- it is about seeing how everything connects.

Interactive graph Color-coded nodes by type -- sources, references, entities, synths. Physics-based layout that animates new nodes into position without restarting the simulation.
Real-time updates The graph updates live as your archive changes. Process new files and watch them appear.
Archive explorer Browse, search, and filter your archive. View any node's summaries, claims, metadata, and relationships. List mode with sorting, filtering, and pagination.
Chat interface Streaming conversational AI with visible thinking, tool calls, and grounded answers. Side-by-side with the graph for visual context.
Agent session replay Review any past agent session -- Seeker searches, Oracle explorations, Weaver operations -- with full telemetry.
The Forge Portal Everything in one interface -- chat, search, graph, jobs, configuration. Accessible from any browser.

Technical detail

Two persistent Sigma.js v3 WebGL renderers using graphology. ForceAtlas2 physics layout. Diff-based graph updates via SignalR. Incremental layout for smooth additions without restarting the simulation.

Why it matters:

Understanding your knowledge is not just about text results -- it is about seeing how everything connects and exploring relationships visually.