RagLeap Core is the open-source engine behind RagLeap — a self-hosted, agentic system that runs your business from your own documents on your own server, with no vendor lock-in.

46 role-based AI Employees: 9 core generalist roles (AI Manager, Secretary, CEO, Sales, Support, HR, Finance, Marketing, Operations) plus 37 vertical-specific global roles (Recruiter, Real Estate Agent, Legal Intake, Healthcare Intake, Insurance Agent, and more) — full list in core/employees/defaults.py.

What it does:

• Self-learning, outcome-weighted memory
• Auto-trigger workflows and escalation
• Full / semi autonomy modes
• Think-act-decide loop, not just retrieval

Quickstart · Docs · Website · Hosted Version · Packages

---

Not to be confused with install.ragleap.com — that's a separate, paid, license-gated self-hosted product (Free tier with a license key, up to Enterprise). ragleap-core (this repo) is MIT-licensed, completely free, and never requires a license key. If you cloned this repo, you're in the right place for a genuinely free, open-source RAG engine.

Install

pip install ragleap-rag

⭐ If this helps you, please consider starring the repo — it genuinely helps more people find it.

Add ragleap-graph too if you want Neo4j-backed knowledge graph retrieval:

pip install ragleap-rag ragleap-graph

Add ragleap-vectorstores too if you want pluggable vector backends beyond ragleap-rag's built-in six (Chroma today):

pip install ragleap-rag ragleap-vectorstores[chroma]
# or, with uv
uv add ragleap-rag ragleap-vectorstores[chroma]

Deploying to Kubernetes? ragleap-ops ships live-tested manifests and a Helm chart for the full stack:

pip install ragleap-ops
# or, with uv
uv add ragleap-ops

Need a generic, reusable chart for your own app (not RagLeap-specific)? ragleap-app-chart takes an arbitrary services: list, not hardcoded names:

pip install ragleap-app-chart
# or, with uv
uv add ragleap-app-chart

Or run the full self-hosted app (channels, web chat UI, Docker Compose) — see Quickstart below. Browse every package at packages.ragleap.com. Try it hands-on with the runnable scripts in examples/ — 01_ingest_and_query.py (upload a document, ask a question via the API) and 02_test_channel_directly.py (test channel answering logic without real bot credentials).

If a RAG chatbot answers questions, RagLeap runs your business

Most open-source RAG projects give you a toolkit — you still have to build the app, wire up a UI, add memory, and connect every channel yourself. RagLeap Core gives you a working chat engine out of the box, and the full RagLeap platform turns it into an AI that actually operates a business.

Without RagLeap	With RagLeap
❌ A different bot for your website, WhatsApp, and Telegram — none of them share memory	✅ One AI across every channel, with memory that persists between them
❌ Your RAG chatbot forgets everything the moment a session ends	✅ Persistent memory — facts and preferences carry across sessions and channels
❌ You're a developer, so you can wire up LangChain — but your team can't manage it	✅ A real dashboard for non-technical owners: settings, analytics, team, billing
❌ Answering customer questions and running the business are two separate systems	✅ Manager AI — an executive assistant that can see analytics, send emails, and manage settings by conversation
❌ Adding a phone line means integrating Twilio, STT, and TTS yourself	✅ Voice AI is built in — real inbound calls, answered and routed automatically
❌ Automating a workflow means writing custom code per integration	✅ n8n workflow automation triggered directly from any conversation

What makes RagLeap Core specifically different

This repo isn't a general-purpose RAG framework you assemble into something — it's the real, working engine that already powers a production AI business platform (see What's in the hosted version below). The code here is honest about being early, but it's extracted from something that already works in the real world, not built as a demo.

Why RagLeap exists

Open-source AI agent projects like OpenClaw took off for a specific reason: people wanted an assistant that runs on their own infrastructure, with their own keys, answering from the chat apps they already use — not a black box hosted by someone else. That same principle is what RagLeap Core is built on for business AI specifically.

Your keys, your infrastructure, your data. RagLeap Core never asks for a system API key. You bring your own Gemini key, you run your own PostgreSQL database, your documents never leave your server unless you choose the hosted version.

Chat is the interface, not a separate dashboard you have to learn. The same way OpenClaw meets people on WhatsApp, Telegram, and Slack, RagLeap's full platform meets business owners on the channels they already use — WhatsApp, Telegram, Discord, and real phone calls — not a new app they have to check.

A real, working system — not an abstract framework. This isn't a toolkit like LangChain where you assemble your own app from primitives. RagLeap Core is the actual chunking → embedding → retrieval → generation pipeline extracted from a production system that already answers real customer questions, at a company that already runs on it.

Built in public, honestly. This repo says clearly what's done and what isn't. No inflated claims, no vaporware Quickstart commands that don't work yet — the Roadmap reflects the real state of the code, updated as it progresses.

What RagLeap Core is

RagLeap Core is a document-grounded chat engine. Upload your documents, ask questions, get cited answers — self-hosted, on your own infrastructure, with your own API key.

WhatsApp, Telegram, and Discord bots are included in this repo too — single-tenant, .env-configured channel adapters that answer from the same document knowledge base. It is the foundation of RagLeap, a hosted AI business manager that adds Voice calling, multi-tenancy, a persistent memory system, and an executive-assistant layer on top of this same core engine.

If RagLeap (hosted) is the business, RagLeap Core is the engine room.

RagLeap Core is right for you if

• ✅ You want a self-hosted RAG chatbot with full control over your data
• ✅ You want to understand exactly how document retrieval and citation works, not use a black box
• ✅ You're comfortable running your own server and your own AI provider key
• ✅ You want to contribute to or extend an open document-QA engine
• ✅ You'd rather see the code than trust a vendor's word on data privacy

What RagLeap Core is not

It's not...	It is...
A hosted product	Self-hosted software you run yourself
Multi-tenant, with persistent cross-session memory	Single-tenant — one bot, one document set, per deployment
A multi-tenant platform	WhatsApp/Telegram/Discord/Voice channel adapters included, single-tenant — multi-tenant routing lives in the hosted version
A no-code SaaS dashboard	A codebase you deploy and configure
Feature-complete with the hosted version	The foundational subset — see Roadmap

Features

📄 Document ingestion	Upload PDFs, text, and common document formats
🔍 RAG retrieval	Vector search over your documents via pgvector
💬 Chat with citations	Answers reference the source document, not a black box
🔌 Bring your own AI key	OpenAI, Gemini, Anthropic, or any OpenAI-compatible endpoint
🌐 Web chat widget	Embed a chat widget on any website
🐳 Docker-based setup	One-command local deployment
🕸️ Knowledge Graph (Neo4j)	Entity extraction and graph-boosted retrieval alongside vector search
🌍 Language detection	Auto-detects document and query language, applied across every channel
🔗 Integrations	Connect MySQL, PostgreSQL, MongoDB, REST APIs, Salesforce, HubSpot, Shopify, Google Sheets, Stripe
🔀 Hybrid search	Combines dense (vector) and sparse (full-text) retrieval via Reciprocal Rank Fusion
⚡ Streaming responses	Answers stream token-by-token instead of waiting for the full response
🔁 Provider fallback	Automatically retries with a backup LLM provider if the primary fails
💰 Token usage reporting	Real per-call token counts from the provider, plus context-size budget trimming
🧑‍💼 AI Employees	Role-based agents (46 default roles) with persistent business-context memory, wired into /chat via role=<role>
🔗 n8n workflow automation	Fire a webhook after the AI replies on WhatsApp/Telegram/Discord — no-code automations triggered directly from a conversation

Architecture

RagLeap Core is the foundation layer of the full RagLeap platform. Here's how it fits into the bigger picture:

flowchart TD
    subgraph Hosted["RagLeap — Hosted Platform (locked)"]
        H1["Manager AI"]
        H2["Multi-tenant AI Employees + Manager AI integration"]
        H3["Persistent Memory (cross-channel, cross-session)"]
        H4["Multi-tenant Billing, Teams & Permissions"]
        H5["Audit History / Compliance logging"]
        H6["Embed Widget Control Center (white-label)"]
        H7["Managed hosting, backups, SLA, support"]
    end

    subgraph Core["RagLeap Core — this repo (open)"]
        WebUI["Web Chat UI"] --> ChatAPI["Chat API"]
        ChatAPI --> WA["WhatsApp"]
        ChatAPI --> TG["Telegram"]
        ChatAPI --> DC["Discord"]
        ChatAPI --> VC["Voice"]

        WA --> N8N["n8n Workflow Trigger (fires after AI reply)"]
        TG --> N8N
        DC --> N8N

        Employees["AI Employees (role context, learned memory)"] --> Provider

        WA --> Ingest["Document Ingest"]
        TG --> Ingest
        DC --> Ingest
        VC --> Ingest

        WA --> RAG["RAG Retrieve"]
        TG --> RAG
        DC --> RAG
        VC --> RAG

        WA --> Provider["AI Provider Adapter"]
        TG --> Provider
        DC --> Provider
        VC --> Provider

        Ingest --> PG[("PostgreSQL + pgvector")]
        RAG --> PG
        Provider --> PG
        PG --> Neo[("Neo4j (Knowledge Graph)")]
    end

    Core -. built on top of .-> Hosted

[locked] = commercial/hosted-only feature, not included in this repository. See below for the full breakdown.

Repo structure

ragleap-core/
├── core/                  # RAG engine — chunking, embedding, retrieval, generation
│   ├── chunker.py
│   ├── embedding.py       # Gemini embeddings (gemini-embedding-001, 3072-dim)
│   ├── retrieval.py       # pgvector cosine search
│   ├── generation.py      # 19-provider BYOK generation (Gemini, OpenAI, Anthropic, etc.)
│   ├── ingest.py          # chunk -> embed -> store pipeline
│   ├── parsers.py         # PDF/DOCX/TXT text extraction
│   ├── employees/         # AI Employees — roles, business profile, learned memory
│   ├── workflows.py       # n8n workflow automation — webhook triggers
│   └── api.py             # FastAPI app — /health, /upload, /chat, /profile, /employees, /n8n-workflows, /webhook/*
├── channels/              # Messaging + voice channel adapters
│   ├── whatsapp/          # Twilio + Gupshup
│   ├── telegram/
│   ├── discord/
│   └── voice/             # Twilio Media Streams, WebSocket server
├── db/
│   └── schema.sql         # documents + chunks tables, pgvector index
├── examples/              # Runnable example scripts
├── .github/workflows/     # CI: compile check, Docker build, smoke tests
├── docker-compose.yml     # app + db + voice services
└── Dockerfile

What's in the hosted version (ragleap.com)

RagLeap Core covers document upload, retrieval, and web chat. The hosted platform builds a full AI business manager on top of it:

Area	What it adds
Manager AI	A private executive assistant for the owner — sees documents, analytics, team permissions, and database connections; can send emails, generate reports, and manage settings by conversation, reachable via Web, WhatsApp, Telegram, or phone call
AI Employees	Single-tenant runtime (46 roles, pgvector-backed learned memory) is open in this repo's core/employees/; the hosted platform adds multi-tenant per-workspace seeding and Manager AI integration on top
Voice AI	Real inbound phone calls via Twilio — speech-to-text, RAG-grounded response, text-to-speech, with owner vs. customer call routing
Multi-channel bots (multi-tenant)	WhatsApp (Twilio or Gupshup), Telegram, and Discord — single-tenant versions are in this repo; the hosted version adds multi-tenancy, per-workspace routing, and shared config across channels
Persistent Memory	Facts and preferences that persist across sessions and channels, not just within a single conversation
Advanced AI Settings	Model selection (Gemini/OpenAI/Claude), temperature tuning, bring-your-own-key per provider, and automatic failover across a fallback key pool
Team Chat	Internal team messaging board per workspace, separate from customer-facing AI chat
n8n Workflows	Single-tenant webhook triggers (WhatsApp/Telegram/Discord) are open in this repo's core/workflows.py; the hosted platform adds multi-tenant per-workspace routing and Voice channel coverage
222+ Languages	This repo includes language detection (langdetect, ~55 languages) across all channels; the hosted platform extends this to 222+ languages with per-user persisted preferences
Integrations & Database Connectors	This repo includes 9 raw connectors (MySQL, PostgreSQL, MongoDB, REST API, Salesforce, HubSpot, Shopify, Google Sheets, Stripe) with on-demand sync; the hosted platform adds AI-suggested automations per channel and developer-level custom automation workflows on top
Analytics Dashboard	Per-provider usage breakdown (OpenAI, Gemini, Claude), query volume, token costs, and daily trends
Team & Billing	Multi-tenant workspaces, team member permissions, subscription plans, usage-based billing
Audit History	Full log of configuration changes — who changed what, and when
Embed Control Center	White-label widget builder — Bubble, Fixed Panel, or Full Page embeds for any website
Data Visibility Controls	Per-document and per-database Public/Private settings — control exactly what customer-facing bots see vs. what's reserved for Manager AI only
Observability & Hallucination Detection	Built-in monitoring layer that audits AI responses for accuracy and flags potential hallucinations
Cloud or Self-Hosted deployment	Run the full platform as managed cloud (subscription) or self-hosted on your own server (one-time license)
Managed hosting	Backups, uptime SLA, and support — zero infrastructure to maintain

This is the standard open-core model — the same approach used by projects like n8n, Supabase, and Cal.com: the engine is free and open, the managed/extended product is commercial.

Full Technical Architecture (Hosted Platform)

This section documents the real internals of the hosted RagLeap platform — gathered by reading actual production source, not summarized from memory or marketing copy. Where something is confirmed not to be live (dead code, an unwired tool), it's labeled as such rather than omitted. This is a living section — deeper subsystems (billing, multi-tenant workspace routing) are deliberately excluded here since they're operational/SaaS infrastructure, not differentiating technology.

System Overview

flowchart TD
    subgraph Channels["Entry points"]
        Web[Web Chat] --- WA[WhatsApp] --- TG[Telegram] --- DC[Discord] --- Vc[Voice / Twilio]
    end

    Channels --> Router{"Owner or customer?<br/>(Voice: verified mobile number match.<br/>Text: personal bot vs. customer-facing config)"}

    Router -->|Owner| MgrL1["Manager AI - Layer 1: Agent Framework<br/>regex router first, LLM classifier only on miss<br/>6 specialist agents, think/act/verify/heal loop"]
    MgrL1 --> MgrL2["Layer 2: Autonomous Loop<br/>off / semi / full modes, per-workspace<br/>action + channel allowlists, approval protocol"]
    MgrL2 --> MgrL3["Layer 3: LLM<br/>only reached when Layer 1 can't match an intent<br/>always given real, current workspace data"]
    MgrL3 -.->|not yet wired, see Known Gaps| RAGShared["Customer-facing RAG index"]

    Router -->|Customer| Addon["Addon DB action check<br/>RealTimeExternalDataService<br/>owner-configured SELECT/UPDATE/INSERT/DELETE"]
    Addon --> RAG["RAG Retrieval<br/>pgvector cosine search + Neo4j entity-graph boost"]
    RAGShared --- RAG

    RAG --> Employees["AI Employees<br/>46 roles, pgvector learned memory,<br/>skill-based context injection"]
    Employees --> Gen["Generation<br/>19-provider BYOK (Gemini/OpenAI/Anthropic/etc.)"]

    Gen --> Mem["Persistent Memory<br/>transactional outbox: embedding/graph/TTS writes<br/>queued in the same DB transaction as the memory row"]

    Ingest["Document / URL Ingestion<br/>OCR fallback, language detection,<br/>canary QA retrieval check before marking complete"] --> RAG

    style MgrL3 fill:#2a2a3a,stroke:#66a
    style RAGShared fill:#3a3a2a,stroke:#aa4

How to read this: solid arrows are confirmed, live data flow. Dotted arrows mark a real gap between two systems that look like they should already be connected but aren't (see RAGQueryTool in the Manager AI section below — Manager AI cannot currently query the customer-facing RAG index despite a tool existing for exactly that purpose). Each subsystem below is documented on its own with the same standard: read from real source, verified live or explicitly marked otherwise.

Manager AI — Three-Layer Design

Manager AI answers most owner requests without calling an LLM at all. A regex-based router matches intent first; only genuinely ambiguous requests reach a model, and even then the model is always given real, current workspace data rather than reasoning blind.

flowchart TD
    Owner["Owner message (Web, Telegram, WhatsApp, Discord, Voice)"] --> Guard["Guardrails: input-length cap, prompt-injection regex"]
    Guard --> Router["RouterAgent"]

    Router -->|"Layer 1: regex match (instant, free)"| Domain{"Domain matched?"}
    Router -->|"Layer 2: tiny 1-token LLM classify (only on regex miss)"| Domain

    Domain -->|email| EmailAgent["EmailAgent"]
    Domain -->|channels| ChannelAgent["ChannelAgent"]
    Domain -->|business| BusinessAgent["BusinessAgent"]
    Domain -->|healing| HealingAgent["HealingAgent"]
    Domain -->|self| SelfAgent["SelfAgent"]
    Domain -->|email_status| EmailStatusAgent["EmailStatusAgent"]
    Domain -->|none| LLMFallback["Full LLM response (real data injected first)"]

    EmailAgent --> Think["agent.think() — deterministic plan, no LLM"]
    ChannelAgent --> Think
    BusinessAgent --> Think
    HealingAgent --> Think
    SelfAgent --> Think
    EmailStatusAgent --> Think

    Think --> Act["agent.act() — executes real Tool classes"]
    Act --> Verify["agent.verify() — did it actually work?"]
    Verify -->|no| Heal["agent.heal() — pattern-matched recovery message"]
    Verify -->|yes| Reply["Reply to owner"]
    Heal -->|still failing| Adapt["Re-plan, retry up to 2x with error context"]
    Adapt --> Think
    Heal -->|healed| Reply

    Act -.->|starts_flow signal| FlowSM["FlowStateMachine — multi-step setup wizard (Telegram/WhatsApp/Gmail connect)"]
    FlowSM --> Reply

    Act --> Obs["observability.py — logs every LLM/tool call/guardrail-block/hallucination to AgentTrace"]
    LLMFallback --> OutGuard["Guardrails: hallucination check (non-blocking, logs only)"]
    OutGuard --> Reply

Verified live. Every node above was confirmed by reading the actual source: api/agent_framework.py (base Agent/Tool/AgentOrchestrator classes), api/agents/*.py (6 specialist agents), api/agents/router_agent.py, api/agents/agent_brain.py, api/agent_state_machine.py, api/guardrails.py, api/observability.py.

Confirmed dead code — not live, listed here so nobody rediscovers them by accident:

• 8 tool classes in api/agent_tools.py are fully implemented but called from nowhere in the codebase: RAGQueryTool, DocumentReadTool, DocumentUploadTool, MemoryReadTool, MemoryWriteTool, MemorySearchTool, AgentStateReadTool, AgentStateWriteTool. Notably, RAGQueryTool means Manager AI does not currently share live query access with the customer-facing RAG system, despite a tool existing for exactly that purpose.
• send_owner_whatsapp, send_owner_telegram, send_owner_sms are each defined twice in api/manager_actions.py. ACTION_REGISTRY (the real dispatch table) is built before the second definitions appear, so it's permanently bound to the first, shorter versions — the second, longer versions are unreachable dead code, unless something imports them directly by name (confirmed: nothing currently does).

Manager AI — Layer 2: Autonomous Loop

Beyond responding to owner messages, Manager AI can act on its own — within limits the owner explicitly configures. Three modes, per workspace: off (owner-initiated only, the default), semi (AI proposes, owner approves via a reply), full (AI executes directly and reports).

flowchart TD
    Trigger["Autonomous trigger (e.g. EmailAgent.scan_and_plan(), a scheduled follow-up)"] --> Dispatch["execute_or_request()"]
    Dispatch --> ModeCheck{"Mode? Action/channel allowlisted?"}
    ModeCheck -->|"off, or not allowlisted"| Skip["Skipped"]
    ModeCheck -->|full| Execute["Execute immediately via execute_fn()"]
    Execute --> Log["log_autonomous_action() — bounded to last 200 entries in memory.knowledge"]
    ModeCheck -->|semi| Pending["Store pending action, keyed by an 8-char action_id"]
    Pending --> RequestApproval["request_approval() — sends 'Reply YES/NO {action_id}' to the owner's configured approval channel"]
    RequestApproval --> OwnerReply{"Owner replies"}
    OwnerReply -->|"YES {action_id}"| ExecutePending["_execute_pending_action() — dispatches to email/whatsapp/telegram/sms/voice"]
    OwnerReply -->|"NO {action_id}"| Reject["Rejected — logged, cancelled, never executed"]
    ExecutePending --> Log
    Reject --> Log
    Log --> DailyReport["generate_autonomy_daily_report() — daily summary of actions taken, grouped by type, failures flagged"]

Verified live, read in full from api/autonomy_engine.py. Two independent gates apply before mode even matters: an actions allowlist and a channels allowlist, both configurable per-workspace — so "full autonomy" doesn't mean unrestricted, it means unrestricted within whatever the owner explicitly turned on. The approval flow in semi mode is a genuine two-way protocol, not a one-shot notification: the owner's exact reply (YES ABC123 / NO ABC123) is parsed and matched back to the specific pending action before anything executes.

Action Dispatch — 70 named actions, one registry

api/manager_actions.py's execute_action(action_type, workspace, params, memory) dispatches by string name through ACTION_REGISTRY, a dict of ~70 real handler functions. Natural-language flexibility comes from deliberate many-to-one aliasing — e.g. get_all_settings is reachable via 5 different phrasings (show_settings, current_settings, ai_settings_info, view_settings), each mapped to the same function — not fuzzy matching.

Two real behaviors worth noting precisely:

• Learns from correction: before executing a channel-config action, execute_action() checks memory.preferences['providers_rejected'] — if the owner previously rejected a provider, it won't silently reconfigure it again.
• Owner vs. customer channels are genuinely separate action families: whatsapp_config/telegram_config/discord_config set up the customer-facing bot; telegram_personal_bot_config/whatsapp_personal_bot_config/discord_personal_bot_config (each a lambda wrapping save_personal_bot_config(ws, platform, ...)) set up the owner's own channel for talking to Manager AI. Same channel types, two distinct configurations.

Document Ingestion Pipeline

flowchart TD
    Upload["Upload (file or URL)"] --> Create["Create Document record + UploadProgress tracker"]
    Create --> Save["Save file to disk"]
    Save --> Parse["file_parser.parse_file()"]

    Parse -->|zip| ZipExpand["Expand: each archive entry becomes its own Document, independently chunked + embedded"]
    Parse -->|other| OCRCheck{"PDF and parsed text too short?"}

    OCRCheck -->|yes| OCR["ocr_pdf() pre-pass — keeps OCR text only if longer/better than parsed"]
    OCRCheck -->|no| LangDetect
    OCR --> LangDetect["Language detection"]

    LangDetect -->|confidence below threshold| Flag["Fall back to workspace default language, flag document for manual review"]
    LangDetect -->|confident| Chunk

    Flag --> Chunk["Chunk text"]
    Chunk -->|custom chunk_size/overlap given| BasicChunker["Basic chunker"]
    Chunk -->|default| DocAwareChunker["DocumentAwareChunker — section-aware"]

    BasicChunker --> Embed["Generate embeddings per chunk"]
    DocAwareChunker --> Embed

    Embed --> Graph{"Neo4j graph_service available?"}
    Graph -->|yes| GraphIndex["Extract entities, upsert document graph"]
    Graph -->|no| SkipGraph["Skip gracefully — logged as warning, not a failure"]

    GraphIndex --> QA
    SkipGraph --> QA["Ingestion QA: canary check"]

    QA --> Canary["Extract top 3 highest-frequency terms from source text, run REAL retrieval queries for each through EnhancedRetrievalService, confirm this document is actually retrievable"]
    Canary -->|hits below threshold| Warn["Flag document with qa_failed warning (hard-fail is opt-in per deployment)"]
    Canary -->|passes| Complete["Mark Document completed"]
    Warn --> Complete

Verified live, read in full from ingestion/pipeline.py (854 lines). The canary QA step is the most distinctive piece: rather than just checking "did the embedding API call succeed," it runs the document's own most distinctive terms back through the real production retrieval path and confirms the document itself shows up in results — proving end-to-end searchability, not just successful ingestion.

Known limitation, relative to the open-source ragleap-core repo: the open-source core/chunker.py only has the basic chunker — DocumentAwareChunker's section-aware chunking is hosted-only.

URL Extraction (documents from web pages and YouTube)

flowchart TD
    URL["Submitted URL"] --> Detect{"YouTube URL pattern?"}

    Detect -->|yes| YT1["Try manually-created English transcript"]
    YT1 -->|unavailable| YT2["Try auto-generated English transcript"]
    YT2 -->|unavailable| YT3["Try any available language transcript"]
    YT3 --> YTMeta["Fetch title, channel, description via page metadata"]
    YTMeta --> Combine1["Combine transcript + metadata into one document"]

    Detect -->|no| Fetch["Fetch page (cloudscraper if available, else requests) with browser-mimicking headers"]
    Fetch -->|403| Rotate["Rotate user-agent, retry (up to 4 agents)"]
    Rotate --> Fetch
    Fetch --> JSCheck{"Bot-wall / JS-challenge page detected?"}
    JSCheck -->|yes| Playwright["Playwright headless browser: load, scroll in 5 steps to trigger lazy content, re-render"]
    JSCheck -->|no| Extract
    Playwright --> Extract["Extract via ranked content-area selectors (main/article/.content/Wikipedia-specific/etc.), strip ads/nav/cookie-banners, dedupe lines"]
    Extract --> Combine2["Combine title + description + content into one document"]

    Combine1 --> Ingest["Feeds into the same ingestion pipeline as uploaded files"]
    Combine2 --> Ingest

Verified live, read in full from ingestion/url_extractor.py (653 lines). Worth noting plainly: this extractor uses real anti-bot-detection techniques (cloudscraper for Cloudflare bypass, rotating user-agents, browser-mimicking request headers) to reliably extract content from sites that actively try to block automated access — a deliberate, real engineering choice, not incidental.

Persistent Memory

flowchart TD
    Write["Any write to memory (owner instruction, learned interaction, correction)"] --> Entry["MemoryEntry row: pgvector embedding (3072-dim), scope (user/workspace), tags, importance score, retention_policy"]
    Entry -->|same DB transaction| Outbox["OutboxEvent queued: embedding.create, graph.create_node, tts.create, etc."]
    Outbox --> Worker["Background worker processes events async — idempotency key prevents duplicate processing, retries on failure with next_retry_at/attempts tracking"]
    Entry -.->|optional| Connector["Connector reference — memory can be written to the customer's own external storage instead of hosted DB"]

Verified live, read from memory/models.py's MemoryEntry and OutboxEvent models. The transactional outbox pattern is genuinely notable: side effects (embedding generation, graph writes) are queued in the same database transaction as the memory write itself, so a crashed background worker can't cause a memory entry to silently end up without its embedding — the event just waits, retried, until it succeeds.

Voice Channel Routing (Hosted)

flowchart TD
    A[Inbound call to workspace Twilio number] --> B["twilio_voice_incoming<br/>(memory/voice_views.py) - confirmed the only<br/>endpoint provisioned as a number's voice_url"]
    B --> C{Caller number matches<br/>owner's verified mobile?}
    C -->|Yes| D[Manager AI voice endpoint<br/>owner-facing]
    C -->|No| E[twilio_voice_speech<br/>customer-facing]

    E --> F["Addon DB action check<br/>RealTimeExternalDataService.match_and_execute_action<br/>NEW: real business-data lookups now run<br/>before RAG, matching the text channels"]
    F --> G["process_voice_query - same /api/v1/query<br/>endpoint every channel uses. Matched DB<br/>results are folded into the query text"]
    G --> H[synthesize_for_call<br/>twilio_voice_service.py]
    H -.->|"Known gap: audio file hosting is<br/>unimplemented - both branches fall back<br/>to Twilio's built-in TTS voice"| I[Caller hears Twilio's built-in voice]

Verified live, read from memory/voice_views.py and memory/twilio_voice_service.py. Two things worth being direct about:

• Fixed: customer voice calls previously had no path to real business data — a caller asking about an order or appointment could only get a RAG answer from documents, never a live database lookup, unlike WhatsApp/Telegram/Discord. twilio_voice_speech now runs the same RealTimeExternalDataService.match_and_execute_action step already used by the text channels before handing off to RAG, so a match gets folded into the query the AI answers from.
• Still open: ElevenLabs TTS is never actually delivered to callers. synthesize_for_call in twilio_voice_service.py has an unimplemented audio-hosting step (# TODO: Implement audio file hosting/S3 upload) — both the success and fallback branches currently produce the same result, Twilio's built-in voice. A separate, unrelated ElevenLabs helper (api/voice_ai.py::generate_elevenlabs_twiml) does correctly upload synthesized audio to the project's R2/S3 storage, but it is not currently called from any live voice path — fixing this gap means wiring that upload logic into synthesize_for_call itself, not assuming the existing helper is already doing the job.

Quickstart

✅ Status: core pipeline verified working. Ingest -> embed -> retrieve -> generate runs end-to-end via Docker Compose, including a clean fresh-clone test. See the Roadmap for what's next (PDF/DOCX support, alternative BYOK providers).

Fastest way to try it — one command checks Docker, clones the repo, and sets up .env for you:

curl -fsSL https://raw.githubusercontent.com/antonyrag/ragleap-core/main/install.sh | bash

(Windows users: run this in Git Bash, not Command Prompt or PowerShell.)

The script will pause after cloning and ask you to add your Gemini API key to .env — get a free one at aistudio.google.com/apikey, then re-run the same command.

Or, the manual way — better if you want to read the code before running anything:

git clone https://github.com/antonyrag/ragleap-core.git
cd ragleap-core
cp .env.example .env
# add your Gemini API key to .env
docker compose up --build -d

Requirements: Docker, Docker Compose, an API key from OpenAI, Google Gemini, or Anthropic.

Try it in 30 seconds — with the stack running, see examples/ for two verified, runnable scripts:

• examples/01_ingest_and_query.py — upload a document and ask a question via the API
• examples/02_test_channel_directly.py — test the WhatsApp/Telegram/Discord answering logic without real bot credentials

Just want the RAG engine as a Python library?

If you don't need the full Docker app — WhatsApp/Telegram/Discord/Voice adapters, the web chat UI, all of it — the core retrieval engine is also published as standalone, pip-installable Python packages:

pip install --index-url https://packages.ragleap.com/simple/ ragleap-rag

• ragleap-rag — the chunking → embedding → retrieval → generation pipeline as a library. Pluggable embeddings (12+ providers), 6 vector backends (FAISS, PgVector, Pinecone, Weaviate, Qdrant, Milvus), cross-encoder reranking, and more.
• ragleap-graph — Neo4j-backed knowledge graph retrieval, usable standalone or alongside ragleap-rag.
• ragleap-vectorstores — pluggable vector backends beyond ragleap-rag core's 6. First backend: Chroma, embedded/local via chromadb's PersistentClient — no server required. Install with pip install ragleap-vectorstores[chroma] or uv add ragleap-vectorstores[chroma].
• ragleap-ops — Kubernetes deployment manifests for RagLeap Core, live-tested end-to-end on a real cluster. Install with pip install ragleap-ops or uv add ragleap-ops.
• ragleap-app-chart — generic, reusable Helm chart for deploying arbitrary services to Kubernetes, not RagLeap-specific. Point it at your own app via a services: list. Install with pip install ragleap-app-chart or uv add ragleap-app-chart.

All five are MIT licensed. Browse the full package index at packages.ragleap.com.

Supported LLM Providers (BYOK)

RagLeap Core is bring-your-own-key only there is no system-provided key for any provider. Set LLM_PROVIDER in .env to choose which one to use for the generation (chat) step. Embeddings currently always use Gemini (gemini-embedding-001), regardless of LLM_PROVIDER.

LLM_PROVIDER value	Required env vars	Notes
gemini (default)	GEMINI_API_KEY	Get a key at aistudio.google.com/apikey
anthropic	ANTHROPIC_API_KEY, ANTHROPIC_MODEL (optional)	Get a key at console.anthropic.com
openai	OPENAI_API_KEY, OPENAI_MODEL
mistral	MISTRAL_API_KEY, MISTRAL_MODEL
groq	GROQ_API_KEY, GROQ_MODEL	Free tier available
together	TOGETHER_API_KEY, TOGETHER_MODEL
openrouter	OPENROUTER_API_KEY, OPENROUTER_MODEL
ollama	OLLAMA_MODEL (no API key needed)	Self-hosted; requires Ollama running locally
deepseek	DEEPSEEK_API_KEY, DEEPSEEK_MODEL
xai	XAI_API_KEY, XAI_MODEL
cohere	COHERE_API_KEY, COHERE_MODEL
perplexity	PERPLEXITY_API_KEY, PERPLEXITY_MODEL
qwen	QWEN_API_KEY, QWEN_MODEL
moonshot	MOONSHOT_API_KEY, MOONSHOT_MODEL
zhipu	ZHIPU_API_KEY, ZHIPU_MODEL
yi	YI_API_KEY, YI_MODEL
baidu	BAIDU_API_KEY, BAIDU_MODEL
minimax	MINIMAX_API_KEY, MINIMAX_MODEL
custom	CUSTOM_API_KEY, CUSTOM_MODEL, CUSTOM_BASE_URL	Any OpenAI-compatible endpoint

Example, switching to Groq in .env:

LLM_PROVIDER=groq
GROQ_API_KEY=your-groq-key
GROQ_MODEL=llama-3.3-70b-versatile

Voice Channel (Twilio)

RagLeap Core includes a real-time voice channel: Twilio Media Streams connects
via WebSocket, your speech is transcribed with OpenAI Whisper, answered by the
core RAG pipeline, and spoken back with OpenAI TTS. The voice-activity
detection and echo-suppression logic is carried over from a production
system tuned against real call traffic.

Runs as a separate service on port 8765 (see docker-compose.yml), since
Twilio's real-time audio protocol needs a raw WebSocket server, not an
HTTP route.

Setup:

1. Set OPENAI_API_KEY in .env (used for both Whisper STT and TTS in v1)
2. Optionally set VOICE_BOT_NAME, VOICE_GREETING, VOICE_TTS_VOICE
3. Point a Twilio phone number's <Connect><Stream> TwiML at
   wss://your-domain.com:8765

Honest status: the WebSocket server, Twilio event protocol handling, and
error handling are verified working. The full Whisper/TTS round-trip has
not yet been live-tested end-to-end (requires OpenAI API credits). If you
try it and hit issues, please open one — this is exactly the kind of
real-world testing this project needs.

Known limitations, carried over from production and not yet fixed here:

• Only OpenAI Whisper (STT) and OpenAI TTS are supported in v1 — Deepgram
  and ElevenLabs (multi-language support) are good-first-issue candidates
• Non-English TTS quality varies since OpenAI's TTS voices are English-tuned
• Typical round-trip latency in production was 6-8 seconds

Knowledge Graph (Neo4j)

RagLeap Core builds a lightweight entity co-occurrence graph alongside its
vector index. When you ingest a document, entities (product names, acronyms,
proper nouns) are extracted and linked in Neo4j. When you ask a question,
the same extraction runs on your query, and any documents linked to matching
entities get a similarity boost in retrieval — on top of, not instead of,
normal pgvector search.

Runs as a fourth Docker Compose service on ports 7475/7688 (remapped
from Neo4j's defaults to avoid colliding with another Neo4j instance on the
same host). If Neo4j is unreachable or NEO4J_PASSWORD is unset, the graph
degrades gracefully — retrieval falls back to pure vector search, ingestion
is unaffected.

Setup:

1. Set NEO4J_URI, NEO4J_USER, and NEO4J_PASSWORD in .env (matching
   the NEO4J_AUTH value in docker-compose.yml)
2. Optionally set DOMAIN_TERMS — a comma-separated list of domain-specific
   terms to boost during extraction (e.g. DOMAIN_TERMS=API,SDK,RAG)

Honest status: entity extraction, document graph writes, entity-based
document lookup, and graph-boosted chat retrieval are all verified working
end-to-end, including in CI (fresh build, real ingest, real query, real
graph lookup). The graph boost is currently a simple additive score bump,
not a full weighted re-ranker — a richer hybrid ranking system is a good
next step for anyone who wants to dig in.

Known limitations:

• Entity extraction is regex-based (CamelCase, acronyms, capitalized
  phrases, plus optional domain terms) — not a trained NER model, so it
  will miss some entities and occasionally include noise
• search_related_entities() (multi-hop graph traversal) is implemented
  but not yet wired into the retrieval pipeline — good-first-issue
  candidate for anyone wanting a project

Language Detection

RagLeap Core auto-detects language during document ingestion (per chunk)
and during chat (per query), using the langdetect library plus
script-based heuristics for CJK, Hangul, and Kana text. Since every
channel (WhatsApp, Telegram, Discord, Voice, and the API directly)
routes through the same core chat pipeline, detection applies
consistently everywhere without per-channel wiring.

Setup: works out of the box with no configuration. Optionally set
DEFAULT_LANGUAGE (fallback when detection fails or text is too short),
LANGUAGE_DETECTION_CONFIDENCE_THRESHOLD (default 0.7), and
LANGUAGE_DETECTION_SUPPORTED_LANGUAGES (comma-separated allowlist,
blank = unrestricted).

Honest status: verified working end-to-end — document-level
detection tested at high confidence (0.9999) on a real mixed-language
document, and query-level detection confirmed working via both the API
and CLI.

Known limitations:

• langdetect covers roughly 55 languages — noticeably fewer than the
  hosted platform's 222+, which layers additional detection and
  per-user language preferences on top
• Short queries in closely-related languages can be misdetected (in
  testing, a short French query was detected as Italian) — this is an
  inherent limitation of statistical detection on short text, not
  specific to this port. A good-first-issue candidate for anyone
  wanting to improve short-query accuracy
• Detection is one-way only: RagLeap Core detects the query's language
  and surfaces it, but does not yet steer the AI's response language
  to match — that's a reasonable next step for a contributor

Integrations

flowchart TD
    subgraph Connectors["10 connectors, one shared interface"]
        C1[MySQL] --- C2[PostgreSQL] --- C3[MongoDB]
        C4[REST API] --- C5[Salesforce] --- C6[HubSpot]
        C7[Shopify] --- C8[Google Sheets] --- C9[Stripe]
        C10[CSV Upload]
    end
    Connectors --> Svc["RealTimeExternalDataService<br/>real-time query, 5-min cache, no separate sync step"]

    Owner["Owner configures an action:<br/>trigger phrase or auto-match on SQL/field names<br/>+ a query template (SELECT / UPDATE / INSERT / DELETE)"] --> Svc

    Svc --> Match["match_and_execute_action(workspace_id, user_message, user_identifier)<br/>extracts order_id / email / phone from the message,<br/>substitutes into the template, runs the query"]

    Match --> Chat["Chat channels (WhatsApp/Telegram/Discord/Web)<br/>api/personal_bot_views.py"]
    Match --> Voice["Voice channel<br/>memory/voice_views.py::twilio_voice_speech<br/>wired in 2026-08"]

    Chat --> RAGCtx["Result injected as context<br/>into the RAG prompt"]
    Voice --> RAGCtx
    RAGCtx --> Answer["AI answers with real account/order/appointment<br/>data, not just document knowledge"]

Verified live, read from api/addon_realtime.py's RealTimeExternalDataService. Two things worth being direct about: match_and_execute_action genuinely supports write queries (UPDATE/INSERT/DELETE), not just read-only lookups — owner-configured, so the safety boundary is whatever SQL the owner writes into the template, not something the framework restricts on its own. And until 2026-08, this action-matching step only ran on chat channels; voice calls had no equivalent, which is the gap closed in the Voice Channel Routing diagram above.

RagLeap Core connects to external databases and business tools, syncing
per-user context to personalize RAG responses. Nine connectors are
included: MySQL, PostgreSQL, MongoDB, generic REST APIs, Salesforce,
HubSpot, Shopify, Google Sheets, and Stripe.

Every CRM/SaaS connector uses credentials you provide directly — a
username/password, a private-app token, an admin API token, a
service-account JSON file, or a secret key, depending on the service.
None require registering an OAuth app; nothing here depends on RagLeap
owning any third-party developer account.

Credentials are encrypted at rest (Fernet/AES-128) before being stored.

Setup:

1. Generate an encryption key: python3 -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
2. Set ADDON_ENCRYPTION_KEY in .env to that value
3. Install the SDK for the connector(s) you want (each is optional — see
   requirements.txt)
4. Create a data source: POST /integrations with name, source_type,
   and the relevant credential fields
5. Test it: POST /integrations/{id}/test
6. Sync it: POST /integrations/{id}/sync

Honest status: verified end-to-end against a real public API —
connection testing, syncing, correct identifier-field matching, and
credential encryption (checked as actual ciphertext in the database,
not just assumed) all confirmed working.

Known limitations:

• 9 of the 18 source types listed in the hosted platform's UI have
  real connectors here. CSV Upload, Snowflake, BigQuery, WooCommerce,
  Airtable, Notion, Razorpay, Slack, and Gmail are good-first-issue
  candidates for anyone wanting to add one
• Sync is on-demand only (POST /integrations/{id}/sync) — no
  scheduled background sync yet, though the schema tracks
  sync_interval_minutes for a future Celery-beat-equivalent
• Synced context isn't automatically injected into chat responses yet
  — each channel adapter would need to know its own user's identifier
  first, which is a reasonable next contribution

Retrieval, Generation & Reliability

Beyond the core RAG pipeline, /chat (and the underlying core.chat.ask())
support several controls aimed at production use: retrieval quality,
response latency, provider reliability, and cost.

Hybrid search (dense + sparse). By default, retrieval combines
pgvector cosine similarity with Postgres full-text search (tsvector/
GIN index), fused via Reciprocal Rank Fusion — catching both semantic
matches and exact keyword/identifier matches a pure embedding search can
miss. Pass hybrid=false to use dense-only retrieval instead (cheaper —
one query instead of two).

Streaming. POST /chat/stream streams the answer as it's generated
(text/plain, chunked transfer) instead of waiting for the full response.
Implemented natively per provider (Gemini, Anthropic, and OpenAI-compatible
each have different streaming APIs — all three are real, not one stubbed).

Provider fallback. Set LLM_FALLBACK_PROVIDERS (comma-separated) to
automatically retry with backup providers if the primary fails — a rate
limit, outage, or bad key on your primary provider doesn't have to mean a
failed request. Each fallback needs its own API key configured normally.
Streaming can only fall back before any text has been sent to the
caller — a mid-stream failure surfaces as an error rather than silently
switching providers and confusing the output.

Generation controls. temperature, system_prompt, and max_tokens
are all real per-call parameters (not just env-var defaults) — build your
own agent behavior on top of RagLeap's retrieval without forking the
library.

Token usage & context budget. Every blocking /chat call returns real
token usage (prompt_tokens, completion_tokens, total_tokens) pulled
directly from the provider's response — not an estimate. Retrieved
context is also trimmed to MAX_CONTEXT_CHARS (default 12000, roughly
4 characters per token for English text) before being sent, dropping the
lowest-ranked chunks first, so you're not paying for more context than
necessary. Set MAX_CONTEXT_CHARS=0 to disable trimming.

Honest status: hybrid search's RRF fusion math verified correct
against hand calculation. Streaming verified working end-to-end for the
default provider. Provider fallback verified with a real broken-primary
test — deliberately invalid API key, confirmed fallback to a working
secondary provider with a correct answer. Token usage and context
trimming verified with real numbers: a 3-chunk retrieval trimmed to 1
chunk under a tight budget reduced actual prompt_tokens by 38% on the
same live API.

Known limitations:

• Token usage reporting is not available for streaming responses — each
  provider's streaming API surfaces usage differently, and doing all
  three correctly is separate, not-yet-done work
• MAX_CONTEXT_CHARS is a character-count approximation (~4 chars/token
  for English), not an exact per-provider tokenizer count
• Hybrid search hasn't been benchmarked for actual ranking-quality
  improvement on a multi-document corpus with genuinely conflicting
  dense vs. sparse rankings — only correctness (fusion math, tokenization
  of unusual identifiers) has been verified so far

Roadmap

• Public repository created
• Core RAG engine extracted and cleaned from production codebase
• Standalone Docker Compose setup (no external Django project dependency)
• Document ingestion module (28+ formats, not just PDF/TXT/DOCX)
• Web chat widget
• Bring-your-own-API-key support (19 providers)
• WhatsApp, Telegram, Discord, and Voice channel adapters (single-tenant)
• Knowledge Graph (Neo4j), language detection, database/CRM integrations
• Contribution guide and good-first-issue labels — 8+ issues labeled, with a real external contributor active on #134
• Community Discord

See ROADMAP.md for the full phase-by-phase history.

Contributing

RagLeap Core is working, tested, and open for contributions now. See CONTRIBUTING.md for how to get started, and check the good first issue label for scoped tasks.

Want to see exactly what's being worked on and what's open to claim? Check the Project board — issues are staged as Good First Issue, Ready (Scoped), or Needs Scoping, so you can pick something that matches how much design work you want to do versus just build.

Student, professor, or looking for a capstone/thesis project? See STUDENT_PROJECTS.md for scoped project ideas at starter, semester, and research-grade levels.

Community

• GitHub Issues — bugs and feature requests
• GitHub Discussions — ideas and questions
• ragleap.com — the hosted product

License

MIT © 2026 RagLeap

Troubleshooting

"could not translate host name 'db'" error after a failed docker compose up:
If your first docker compose up attempt fails (e.g. a port conflict on 5433 or 8000), a retry can sometimes leave the database container attached to a stale, orphaned Docker network. Fix:

docker compose down
docker network prune -f
docker compose up --build -d

Port 5433 or 8000 already in use:
Another instance of this project (or something else) is using the port. Either stop it, or change the host-side port mapping in docker-compose.yml (the "5433:5432" and "8000:8000" lines) to something free.

Add AI Employees example scripts to examples/

Documentation reference and guidelines for #135.