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
Agent trace
- Planning
- Knowledge search
- Drafting
Running on a real production stack
How it works
Ingest
Upload PDFs, docs and links. BrainStack reads, chunks and indexes everything into your company's private knowledge base — watch it happen live.
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.
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.
Upload
PDFs, Word docs, URLs
handbook.pdf · 4 pages
Extract
Text + metadata per page
pages → text · tenant tagged
Chunk
Overlapping passages
~400 tokens · overlap kept
Embed
Text becomes vectors
chunks → embeddings, in parallel
Index
Into your private namespace
Pinecone · your tenant only
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.
Question
from anyone on the team
Planner
decides what's needed
Router
picks the tools
Knowledge · Web · Company MCProle-gated
native RAG + web, MCP only for managers
Synthesize
grounded answer + citations
Reflect
grounding critic · max 2 attempts
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
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.
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
2 facts saved to long-term memory
Friday · conversation #2
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.