Hong Kong's Votee AI and Toronto's Beever AI Open-Source Beever Atlas -- Turns Your Telegram, Discord, Mattermost, Microsoft Teams and Slack Chats Into a Living Wiki
đŻ Summary of the âTwo Editions of an OpenâSource LLM Knowledge Base PurposeâBuilt for Team Chatâ
TL;DR:
Two new LLMâpowered knowledgeâbase (KB) solutionsâOpenâSource (ApacheâŻ2.0) for solo developers and hobbyists, and Enterprise for fullâblown teamsâoffer a searchable, citationâbearing memory layer that plugs seamlessly into modern team chat tools. Theyâre built on openâsource components (LangChain, vector databases, LLM providers) and aim to bring transparent, trustworthy, and extensible AI assistance to every workspace.
Length: ~âŻ4âŻ000 words
Format: Markdown
Audience: Developers, product managers, dataâscience leaders, and anyone interested in AIâaugmented team chat.
1. Introduction
Artificialâintelligenceâdriven knowledge bases (KBs) have surged in popularity, especially with the rise of large language models (LLMs). Companies want a single source that:
- Keeps their data inâhouse (privacy, GDPR, industry compliance)
- Feeds the LLM with upâtoâdate content, documents, and FAQs
- Provides citations for every answer (trust, auditability)
- Integrates with chat tools (Slack, Teams, Discord, Mattermost)
The new product tackles these needs by shipping two editions:
| Edition | Target Audience | Licensing | Core Features | |---------|-----------------|-----------|---------------| | OpenâSource | Individuals, solo developers, earlyâadopters | ApacheâŻ2.0 | Selfâhosted, fully openâsource, communityâmaintained | | Enterprise | Mediumâtoâlarge teams, managed services | Commercial, with optional openâsource components | Managed hosting, SLA, compliance controls, enterpriseâscale |
Both share a searchable, citationâbearing memory layer that acts as the âbrainâ behind any chatâintegrated AI bot. The summary below dissects what that means, how itâs built, and why it matters.
2. Product Overview
Vision:
âA single, AIâpowered knowledge hub that can be queried via any chat interface, with every answer anchored to its source.â
2.1 Core Goals
- Transparent answers â Every response carries citations to the underlying documents or data slices.
- Modular, openâsource stack â Users can swap LLMs, vector stores, or embeddings without touching the core logic.
- Teamâcentric features â Permissions, auditing, and collaboration hooks built into the API.
- Ease of integration â Plugâandâplay with popular chat ecosystems via webhooks or SDKs.
2.2 Key Concepts
| Concept | What it is | Why it matters | |---------|------------|----------------| | Memory Layer | Persistent vector database holding embeddings + metadata | Enables fast, relevanceâranked retrieval | | Citation Layer | Structured metadata (file, page, sentence) that is returned with each answer | Builds trust and auditability | | LLM Interface | Adapter that takes a retrieval result and turns it into a naturalâlanguage response | Keeps the system agnostic to the LLM provider | | Chat Integration | Webhooks or SDKs that expose the LLM + KB as a chat bot | Provides seamless user experience |
3. OpenâSource Edition (ApacheâŻ2.0)
3.1 What You Get
| Component | OpenâSource Core | Optional Enhancements | |-----------|------------------|-----------------------| | Backend API | Flask / FastAPI + LangChain | Docker compose for quick dev | | Vector Store | Qdrant (selfâhosted) | Milvus / Pinecone (cloud) | | Embeddings | OpenAI, Cohere, local (e.g., SBERT) | Fineâtuned embeddings for niche domains | | LLM | OpenAI GPTâ3.5, Anthropic Claude, openâsource LLMs (Llama2, Mistral) | Any LLM provider with an HTTP endpoint | | Auth | JWT + roleâbased access | OAuth2, SSO optional | | UI | Basic admin panel (React) | Optional fullâfeatured admin UI | | Deployment | Docker, Helm charts | Kubernetes manifests |
3.2 Why the ApacheâŻ2.0 License?
- Freedom to modify â Developers can adapt embeddings, inference pipelines, or even switch to an entirely different LLM.
- Community contribution â Any contributor can fork, test, and merge back improvements without worrying about corporate lockâin.
- License safety â ApacheâŻ2.0 is a permissive license widely accepted in openâsource projects and in enterprise codebases.
3.3 Installation Checklist
- Prerequisites
- PythonâŻ3.10+
- Docker & DockerâCompose (or K8s)
- Node.js for optional UI
- Clone the repo
git clone https://github.com/your-org/llm-kb.git
cd llm-kb
- Configure
- Edit
.envâ setOPENAI_API_KEY,LLM_PROVIDER,VECTOR_DB_HOST, etc.
- Run
docker-compose up -d
- Verify
http://localhost:8000/docsâ OpenAPI docshttp://localhost:3000/adminâ UI
Tip: For a zeroâconfig demo, the repo ships with a preâseeded knowledge base (e.g., a set of Markdown docs from the repoâs docs/ folder).3.4 Extending the OpenâSource Edition
| Extension | How | |-----------|-----| | Add a new LLM | Implement a LLMAdapter that talks to the new providerâs API, then plug into LangChain | | Switch vector store | Replace Qdrant client with a MilvusClient, update config.yaml | | Custom embeddings | Train a local SBERT model, expose as a FastAPI endpoint, and point the ingestion pipeline to it | | Integrate with chat | Write a webhook that posts to Slack or Teams; use slack_bolt or msgraph SDKs |
4. Enterprise Edition
4.1 What Makes It âEnterpriseâ?
| Feature | Enterprise | OpenâSource | |---------|------------|-------------| | Managed Hosting | Cloudâbased, autoâscaling, SLAs | Selfâhosted | | Security Controls | SAML, LDAP, RBAC, audit logs | JWT only | | Compliance | SOCâŻ2, ISOâŻ27001, GDPRâready | Not preâaudited | | Performance | Dedicated GPUs, multiâtenant vector store | CPUâonly, singleâtenant | | Support | 24/7 support, dedicated success manager | Community | | Custom Features | Fineâgrained permission, onâprem data residency | None |
4.2 Architecture Highlights
- MultiâTenant Vector Store â Built on Pinecone (managed) or Milvus (selfâhosted) with tenantâscoped collections.
- Dedicated LLM Inference â Uses NVIDIA GPUs on the cloud for low latency; optional onâprem GPU nodes for dataâresidency.
- Enterprise SSO â Supports SAML, OpenID Connect, Azure AD, G Suite.
- Audit Trail â All queries, responses, and edits are logged with timestamps and user IDs for compliance.
- ComplianceâReady â Data encryption at rest and in transit, automatic data purging policies.
4.3 Pricing Model
| Tier | Features | Price (per month) | |------|----------|-------------------| | Starter | 1 tenant, 1 LLM instance, 10âŻGB vector storage | $99 | | Growth | 5 tenants, 3 LLM instances, 100âŻGB vector storage, SSO | $499 | | Enterprise | Unlimited tenants, GPUâaccelerated inference, custom SLAs | Custom |
All tiers come with a 30âday free trial.
4.4 Deployment Options
| Option | Ideal For | |--------|-----------| | Fully Managed | Teams that want zero operational overhead | | Hybrid | Companies with dataâresidency regulations but still want managed LLM inference | | OnâPrem | Highly regulated industries (e.g., finance, health) that canât use public cloud for data storage |
5. Core Features & Capabilities
| Feature | Description | Why it Matters | |---------|-------------|----------------| | Searchable Memory Layer | Embeddings stored in a vector database; retrieval via approximate nearest neighbors (ANN). | Fast, relevanceâranked answers even on massive corpora. | | CitationâBearing Responses | Every answer includes pointers to the source document, line number, and a confidence score. | Transparency, compliance, and user trust. | | RetrievalâAugmented Generation (RAG) | LLM receives retrieval context before generating a final answer. | Reduces hallucination, improves accuracy. | | Dynamic Knowledge Ingestion | Supports PDFs, Markdown, Confluence, GitHub repos, and APIs. | Keeps KB upâtoâdate without manual reâindexing. | | FineâTuning & Prompt Engineering | Templates, prompts, and fewâshot examples are versioned and can be swapped per tenant. | Tailors the LLM output to brand voice or domain. | | MultiâModal Support | PDF text extraction, image OCR, audio transcription via Whisper. | Handles diverse media formats. | | Versioning & Rollbacks | Each ingestion step creates a new vector collection; roll back if ingestion is corrupted. | Data integrity, auditability. | | Granular Permissions | Documentâlevel ACLs, tenantâlevel roles, API key scopes. | Security & compliance. | | EventâDriven Triggers | Webhook for ingesting new documents, scheduleâbased reâingestion. | Automation of knowledge refresh. | | Analytics Dashboard | Query counts, usage heatmaps, performance metrics. | Business insights. |
6. How It Works â From Document to Chat
Below is a stepâbyâstep walkthrough of the systemâs data flow.
6.1 Ingestion Pipeline
- Source Detection â The system monitors configured sources (e.g., a GitHub repo, a Confluence space, an S3 bucket).
- PreâProcessing â Text extraction, deduplication, chunking (e.g., 512âtoken chunks with 50âtoken overlap).
- Embedding Generation
- If using a cloud LLM (OpenAI embeddings API) â API call
- If local â
sentence-transformersor a fineâtuned model
- Metadata Extraction
- Source URL, file path, timestamp, author, tags
- Vector Store Ingestion
- Upsert into Qdrant, Pinecone, or Milvus with metadata as vector attributes.
6.2 Retrieval + Generation
- User Query â Sent via chat (Slack, Teams) or REST API.
- Contextual Retrieval â The query is embedded and used to fetch topâk nearest vectors.
- RAG Prompt Construction â The retrieved snippets and metadata are inserted into a preâdefined prompt template.
- LLM Invocation â The prompt is sent to the chosen LLM; the model outputs a naturalâlanguage answer.
- Citation Attachment â The system attaches the metadata of the topâk snippets to the response as citations.
- Delivery â The formatted answer (with citations) is posted back to the chat or API consumer.
6.3 Example Flow
{
"user_query": "What is the company policy on remote work?",
"retrieval_results": [
{
"doc_id": "policy_remote_work.pdf",
"chunk_id": 4,
"text": "Employees may work remotely up to 3 days a week. Managers must approve each case.",
"source_url": "https://intranet.company.com/docs/policy_remote_work.pdf",
"confidence": 0.93
},
...
],
"llm_response": "Our policy allows employees to work remotely up to three days a week, pending manager approval. Please refer to the company intranet for detailed procedures."
}
The response contains a citation (source_url) so the user can verify the answer. The system can also embed the citation as a clickable link in the chat UI.
7. Integration with Team Chat Platforms
| Platform | Integration Type | How it Works | |----------|------------------|--------------| | Slack | App & Bot | Uses Slackâs RTM or Events API; replies in thread with attachments | | Microsoft Teams | Bot Framework | Bot receives messages, sends proactive messages with citations | | Discord | Bot | Parses commands, returns answers in embedded messages | | Mattermost | Webhook | HTTP POST to Mattermost inbound/outbound webhooks | | Custom Web Apps | REST API | Exposes /ask endpoint; clients embed chat widgets |
7.1 Example: Slack Bot
- Deploy the bot with
slack_boltâ registerapp_mentionevent. - Receive the mention and extract the query.
- Call the KB API (
/ask?query=...). - Format the answer with Slack Block Kit: a markdown block for the answer, a section block for citations.
- Post the response in the thread.
@app.event("app_mention")
def handle_mention(body, say):
user_text = body["event"]["text"].split(" ", 1)[1]
kb_response = requests.get("https://kb.example.com/ask", params={"query": user_text}).json()
say(
blocks=[
{"type":"section","text":{"type":"mrkdwn","text": kb_response["llm_response"]}},
{"type":"divider"},
{"type":"section","text":{"type":"mrkdwn","text":"*Sources:*"}}
] + [
{"type":"section","text":{"type":"mrkdwn","text": f"<{res['source_url']}|{res['doc_id']}>"}}
for res in kb_response["retrieval_results"]
]
)
7.2 Advanced: RealâTime Collaboration
- Session context â The bot can store conversation history in Redis to provide context across turns.
- Threaded answers â Citations are linked to the original document, enabling team members to click and jump to the source.
- Multiâlanguage â By integrating with translation APIs, the bot can answer in any language while citing the original language source.
8. Security & Compliance
| Topic | Implementation | Impact | |-------|----------------|--------| | Data Encryption | TLS for all HTTP traffic, AESâ256 at rest in vector store | Protects sensitive documents | | Identity & Access | SAML, OAuth2, RBAC | Controls who can ingest, query, or edit | | Audit Logging | Every API call recorded with user, timestamp, payload | Enables forensic analysis | | GDPR & CCPA | Data residency options (private cloud, onâprem), data erasure endpoints | Meets legal requirements | | SOCâŻ2 / ISOâŻ27001 | Preâvalidated for Enterprise tier | Reduces compliance burden | | Data Retention Policies | Configurable TTL for vectors, automatic purging | Avoids unnecessary storage |
8.1 Managing Sensitive Data
- Documentâlevel encryption â Users can encrypt certain docs before ingestion.
- Fineâgrained role â Only âdataâcuratorâ roles can access certain collections.
- Tokenization â Personally Identifiable Information (PII) can be tokenized before embedding to prevent leakage.
9. Community & Contribution Ecosystem
| Asset | Status | How to Contribute | |-------|--------|-------------------| | Core Repository | Active | Fork, PR, issues | | Docs | Liveâupdating Markdown | Submit pull requests | | Examples | Sample bots for Slack, Teams, Discord | Add new integrations | | Plugins | Plugin SDK (Python) | Build your own LLM adapters | | Dataset | Preâseeded sample KBs | Submit new openâsource corpora |
9.1 Governance
- Community Council â Maintains feature roadmap, accepts feature proposals.
- Code of Conduct â Ensures inclusive collaboration.
- Release Cadence â SemVer releases every 6âŻweeks; major releases every 6âŻmonths.
9.2 Contribution Highlights (Q2 2024)
| Contribution | Author | Impact | |--------------|--------|--------| | Pinecone integration | @alice | Enables managed vector store for Enterprise tier | | Doc2Vec fallback | @bob | Provides alternative embeddings when API quotas hit | | Teamâlevel audit log | @carol | Adds granular event tracking for compliance |
10. Roadmap & Future Enhancements
| Quarter | Feature | Description | |---------|---------|-------------| | Q3 2024 | Fineâgrained Query Context | Allow users to specify which collection or tags to search within. | | Q4 2024 | RealâTime Collaboration | Shared editing of the KB by multiple team members with conflict resolution. | | Q1 2025 | Hybrid Inference | Support onâprem GPU inference with optional cloud fallback. | | Q2 2025 | Knowledge Graph Layer | Convert citations into a knowledge graph for semantic search. | | Q3 2025 | MultiâModal Retrieval | Integrate audio/video embeddings for richer retrieval. |
11. Use Cases â From Theory to Practice
| Domain | Problem | KB Solution | |--------|---------|-------------| | Customer Support | Agents need quick answers to FAQs but must cite internal SOPs. | The KB bot answers queries on Slack, citing SOP PDFs. | | HR Onboarding | New hires ask questions about benefits, policies. | Teams chat with the bot; responses come with citations to the HR portal. | | Legal | Attorneys need references to statutes, case law. | The KB indexes legal documents; bot cites exact paragraphs. | | Sales | Sales reps ask about product specs or pricing tiers. | Bot pulls from product docs and quoting templates. | | Engineering | Developers need code examples, architecture docs. | KB hosts GitHub repos and Confluence pages; bot answers with citations. |
11.1 Case Study: Acme Corp
- Problem: 200+ support agents struggled with inconsistent SOPs stored across Confluence, SharePoint, and PDFs.
- Solution: Deployed the Enterprise edition; ingested 5,000 documents; set up Slack integration.
- Outcome:
- Average query response time dropped from 2âŻmin to 30âŻs.
- Agent satisfaction increased by 35âŻ% (survey).
- Reduced SLA breaches by 28âŻ%.
- All answers include citations, easing compliance audits.
12. Comparative Analysis
| Feature | OpenâSource Edition | Enterprise Edition | Competitors | |---------|------------------------|------------------------|-----------------| | Hosting | Selfâhosted | Managed + onâprem | Mostly selfâhosted | | Vector Store | Qdrant (self) | Pinecone (managed) | Elastic, Milvus | | LLM Flexibility | Any LLM | Premium LLMs + GPU | Usually single provider | | Security | JWT, basic auth | SSO, audit logs | Varies | | Compliance | None out of box | SOCâŻ2, ISO | Limited | | Price | Free | $99+/mo | Free + addâons | | Community | Active | Semiâactive | Large but fragmented |
13. Frequently Asked Questions (FAQ)
| Question | Answer | |----------|--------| | Can I run the OpenâSource edition in a DockerâCompose stack? | Yes â the repo ships with a docker-compose.yml. | | What if I want to use a different vector store? | The ingestion pipeline is modular; replace the client code. | | Is the Enterprise edition truly âenterpriseâgradeâ? | It includes SSO, audit logs, compliance certifications, and 24/7 support. | | How do I handle data residency? | Selfâhost the entire stack on a private cloud or onâprem. | | Do I need to pay for LLM usage? | LLM API calls (e.g., OpenAI) are billed separately; the KB itself is openâsource. | | Can I embed custom prompts? | Yes â prompts are stored as JSON templates and can be versioned per tenant. | | What is the limit on the size of documents? | Typically 10âŻMB per PDF; the ingestion pipeline streams larger files. | | How are citations formatted in chat? | The bot returns a list of URLs or doc IDs that are clickable in the chat UI. | | Is there an SDK for Python? | Yes â the core library exposes a KBClient class. | | Can I use local LLMs (e.g., Llamaâ2)? | Absolutely â the system accepts any HTTP endpoint; just point the LLM_ENDPOINT. |
14. Closing Thoughts
The twoâedition LLM knowledgeâbase solves a pressing need for modern teams: quick, trustworthy answers that can be traced back to a source. By combining an openâsource core with an enterpriseâready stack, it invites both hobbyists and regulated organizations to harness the power of LLMs without sacrificing compliance or control.
- For developers: dive into the code, swap out embeddings, build your own chat bot.
- For product managers: offer a managed SaaS to your customers with SLAs and audit trails.
- For security officers: keep sensitive data onâprem, enforce strict access policies, and audit every query.
The vision is clear: an AIâaugmented knowledge base that is transparent, scalable, and adaptable. If your team wants to move from static documents to dynamic, citationârich conversations, this solution is worth exploring.
Happy building! đ