The AI knowledge platform for your organization

Ask in plain language, get answers grounded in your own documents — cited, transparent, and able to act in your tools.

No credit card · Isolated per company · Your data never trains a model

app.brainstack.spaceEmployee

Agent trace

  • Planning
  • Knowledge search
  • Drafting

Running on a real production stack

LangGraph agentClaudePineconeFastAPIPostgreSQLSupabaseRedisMCP protocolNext.jsTailwind CSSVercelDocker

How it works

01

Ingest

Upload PDFs, docs and links. BrainStack reads, chunks and indexes everything into your company's private knowledge base — watch it happen live.

02

Ask

Anyone on the team asks in plain language. The agent searches your knowledge (and the web when useful) and streams back a cited, grounded answer.

03

Act

Managers go further: “assign this ticket to Priya” hits your real systems through a standard protocol — with permissions enforced by design.

Knowledge ingestion

Every upload runs a live pipeline — extracted, chunked, embedded and indexed into your company's own private namespace. No black box, no waiting and wondering.

  1. Upload

    PDFs, Word docs, URLs

    handbook.pdf · 4 pages

  2. Extract

    Text + metadata per page

    pages → text · tenant tagged

  3. Chunk

    Overlapping passages

    ~400 tokens · overlap kept

  4. Embed

    Text becomes vectors

    chunks → embeddings, in parallel

  5. Index

    Into your private namespace

    Pinecone · your tenant only

0/312 chunks indexed

Retrieval lets it know.

Your documents become searchable meaning. Every answer is built strictly from what your company actually knows — and every fact links back to its source page.

RAG · embeddings · semantic search · citations

Connections let it do.

Your ticketing, HR and analytics systems plug in over MCP — an open standard. The agent takes real actions in your tools, gated by each person's role.

MCP · tool discovery · role-based access

The brain is one agent. The knowledge and the hands are yours.

No black box

Every answer is a visible plan, not a mystery. The agent decides what it needs, picks its tools, grounds the draft in real sources, and double-checks itself before a single token reaches your screen.

  1. Question

    from anyone on the team

  2. Planner

    decides what's needed

  3. Router

    picks the tools

  4. Knowledge · Web · Company MCProle-gated

    native RAG + web, MCP only for managers

  5. Synthesize

    grounded answer + citations

  6. Reflect

    grounding critic · max 2 attempts

  7. Streamed answer

    token by token, live trace

⟳ self-corrects before answering · hard cap of 2 attempts

This exact trace streams live in the product while the agent works — planning, searching, acting, drafting.

Advanced retrieval

Meaning-based search is brilliant until someone asks about an exact error code or SKU. BrainStack runs dense and keyword retrieval side by side and fuses the rankings — so both kinds of questions land.

Dense retrieval

Pinecone · top-25 by meaning

  • returns & refunds policy0.89
  • expense approval limits0.71
  • client dispute handling0.63
  • ERR_4021 troubleshootingnot in top-25

BM25 keyword

Postgres chunks · top-25 by exact tokens

  • ERR_4021 troubleshootingmatch
  • error code referencematch
  • returns & refunds policymatch
  • on-call escalationmatch

Fused result

RRF merge → rerank* → top-6 to the agent

  • ERR_4021 troubleshooting#1
  • returns & refunds policy#2
  • error code reference#3
  • expense approval limits#4
  • client dispute handling#5
  • on-call escalation#6

Meaning-based search missed the exact codeERR_4021— keyword search caught it, and rank fusion put it first. Every answer gets both.

Evaluation

A demo is a claim. A number is proof.

Every answer BrainStack gives is scored by an automated judge for faithfulness to its sources and relevance to the question. These are the latest production scores — live on the Evaluation page, not a slide.

0.000

Faithfulness

1.000 with the cross-encoder reranker on

0.000

Answer relevance

LLM-as-judge, strict JSON scoring

0.000

Retrieval hit-rate

expected passage found — every time

0.000

Citation validity

every [n] resolves to a real source

22 golden questions2 refusal traps — both refused correctlymeasured against production, not localhostruns stored & browsable on the Evaluation page

The platform

The chat is just the entry point. Everything around it — the library, the trace, the citations, the analytics — is what makes it a product your company can trust.

Intelligence

Grounded Q&A

Answers built only from your own knowledge — with “I don't know” instead of made-up facts.

Intelligence

Live agent trace

Watch every reasoning step in real time: planning, searching, acting, drafting.

Trust

Inline citations

Click any [1] to open the source at the exact page, with the grounding passage beside it.

Knowledge

Multi-source ingestion

PDFs, Word docs, URLs — parsed, chunked and indexed with live progress.

Actions

Role-based actions

Managers assign tickets and pull analytics from chat. Employees can't — by construction.

Insights

Quality you can measure

Faithfulness scores, latency, cost per answer — a dashboard, not a demo.

Platform

True multi-tenancy

Every company's knowledge lives in its own isolated namespace. Always.

Platform

Full observability

Per-request traces of what was retrieved, which tools ran, and what it cost.

Memory

Long conversations compress into summaries instead of falling off a cliff, and durable facts you share are kept as long-term memory — recalled in any future conversation. You stay in control: every remembered fact is visible on the Memory page, and forgetting one is a single click.

  • Short-term

    A sliding window of recent turns, with older context summarized automatically — follow-ups just work.

  • Long-term

    Facts you state are embedded and stored per user, recalled only when they're actually relevant.

Tuesday · conversation #1

I'm Alex — I run the Bangalore office. What's our refund policy?

2 facts saved to long-term memory

days later · a brand-new conversation

Friday · conversation #2

What's my name?
You're Alex — and you run the Bangalore office.

recalled from memory · verified live in production

Security & trust

Multi-tenant AI has a scary failure mode: someone else's confidential data paraphrased into a fluent answer. BrainStack is architected so that can't happen.

Isolated per tenant

Each company's vectors live in their own namespace — cross-company leaks are impossible by construction.

Permissions at the protocol

An employee's agent never even discovers manager tools. Capability-based, not prompt-based.

Your data stays yours

Documents ground answers for your workspace only. Nothing you upload trains a model.

FAQ

Completely. Every workspace's knowledge lives in its own namespace, keyed to your company — a query from one tenant physically cannot reach another tenant's data. It's the first invariant the platform was designed around.

Set up a workspace in minutes. Upload knowledge, invite the team, and start asking.