Knowledge Base

Your business knowledge,
instantly searchable by AI

Upload your knowledge base, policies, and product specs. Reply Flow chunks, embeds, and searches them intelligently — so your agents answer from your material, not the internet.

7-day free trial · Works with documents you already have
Upload & Processing

Drop it in.
We handle
the rest.

Paste text directly or upload files in 30+ formats — from PDFs and DOCX files to spreadsheets and code. Every file goes through a six-step processing pipeline that turns raw content into search-ready knowledge.

1 GB per file limit. Text entries accept a title and pasted content directly — no file upload required.

30+ supported formats
Documents
TXT MD PDF DOCX HTML
Data
CSV TSV JSON JSONL XLSX XLS
Code
.ts .py .go .java .rb .php .rs .swift + 12 more
Processing pipeline
  1. 01
    Classify Detects file type and picks the right extraction path
  2. 02
    Extract Pulls clean text from PDFs, DOCX, spreadsheets, code
  3. 03
    Clean Strips noise, normalises whitespace, fixes encoding
  4. 04
    Chunk Splits into semantically coherent passages
  5. 05
    Embed Generates vector embeddings via OpenAI text-embedding-3-small
  6. 06
    Store Written to Postgres with HNSW + GIN indexes
semantic chunker · live
Source document
returns-policy.md
4 200 words · 23 sections
cosine distance · sentence windows
sentence 1 ▲ split at 95th pct sentence 20
chunks produced
chunk 01
returns-policy.md › Eligibility Window
312 chars · embedded
chunk 02
returns-policy.md › How to Start a Return
284 chars · embedded
chunk 03 …
returns-policy.md › Refund Timelines
331 chars · embedded
Differentiator · Semantic Chunking

Chunks that cut
at meaning,
not character count.

Most systems split at a fixed character limit and call it done. Reply Flow reads the document structure, embeds consecutive sentence windows, and finds where topics actually change — using cosine distance to locate the natural seams in your content.

  • Prose & Markdown Semantic chunking — parses H1–H6 headings, embeds buffered sentence windows, calculates cosine distance between consecutive embeddings, splits at the 95th percentile of distances to detect real topic shifts.
  • Structured data Pre-structured chunking — respects row/record boundaries in CSV, TSV, JSONL, and spreadsheets so a chunk is never half a row.
  • Chunk prefix Every chunk is tagged with the source title and section hierarchy, so the AI always knows what document it's drawing from.
  • Configurable Target chunk size (default 2 000 chars), overlap (200 chars), minimum size (100 chars) — adjustable per knowledge base.
Differentiator · Hybrid Search

Finds what it
means,
not just what
you typed.

Every search query is automatically classified into the right retrieval strategy before a single result is fetched. Short lookups go to full-text. Conceptual questions go to vector. Everything else blends both — merged with Reciprocal Rank Fusion so neither signal drowns the other out.

After results are ranked, Claude Haiku generates a one-sentence relevance explanation for each — so the AI agent understands not just the text, but why it was retrieved.

K
Keyword search (FTS)
When Short queries, identifiers, quoted phrases

PostgreSQL full-text search with a GIN index, English language config

S
Semantic search
When Long questions, conceptual or abstract terms

1 536-dim vector similarity via HNSW index — finds meaning, not just words

H
Hybrid (RRF)
When Mixed queries — most real questions land here

Reciprocal Rank Fusion (k = 60) merges vector + FTS rankings into a single ordered list

Reciprocal Rank Fusion (RRF, k = 60) — a well-studied fusion algorithm that combines ranked lists without needing to tune per-score weights. Vector and full-text results each contribute independent ranking signals; RRF merges them into a final list that outperforms either alone.

Per-Agent Configuration

Different agents.
Different
knowledge.

Not every agent needs to know everything. Attach specific knowledge bases to specific agents or scenarios — and tell the AI how to use each one.

  • Always include KB is injected into every turn. Best for core policies and product specs the agent references constantly.
  • Fallback only KB is searched only when the agent's scenario instructions don't give a clear answer. Reduces token cost for niche material.
  • Per-scenario Override at scenario level — your returns agent references the returns KB; your product agent references the catalogue. Same agent, different knowledge per situation.
  • Custom instructions Write a note per KB: "Treat prices here as current. If in doubt, say 'check with us'." The AI follows the instruction, not just the text.
Knowledge Base — Mira
agent config
Returns & Refunds Policy
42 chunks · last updated 3 days ago
"Always check here before quoting return windows."
always
Product Catalogue Q2 2026
218 chunks · last updated today
"Use pricing from here. Say 'let me check' for custom orders."
fallback
Bulk Pricing Tiers
12 chunks · scenario: bulk-order-pricing
scenario
Entry & Chunk Management

Your knowledge.
Your control.

Drill into individual chunks, edit them in place, or trigger a full re-embed when source material changes. Nothing is a black box.

  • View chunks Browse individual chunks with content and metadata — see exactly what the AI will read.
  • Edit chunks Change any chunk in place. Embeddings update automatically on save.
  • Delete chunks Remove a chunk and the index re-compacts immediately.
  • Re-embed entry Force a full re-embedding of an entire document on demand — useful after source material changes.
  • Full CRUD Create, read, update, delete knowledge base entries and entire knowledge bases.
Returns & Refunds Policy — 42 chunks
01
returns-policy.md › Eligibility Window

Items may be returned within 30 days of delivery provided they are in original condition with tags attached. Digital downloads are non-refundable…

312 chars · 1 536-dim embedding
02
returns-policy.md › How to Start a Return

To initiate a return, log into your account and navigate to Orders. Select the item and click "Return this item". You will receive a prepaid label…

284 chars · 1 536-dim embedding
03
returns-policy.md › Refund Timelines

Approved refunds are processed within 3–5 business days to your original payment method. Bank transfer refunds may take up to 10 days…

331 chars · 1 536-dim embedding
Showing 3 of 42 chunks → View all
What it is · What it isn't
Manual upload

Upload files or paste text. You control exactly what goes into the knowledge base.

Context injection

Relevant chunks are injected into the agent's context before it drafts a reply — grounding answers in your material.

No URL ingestion

There's no "point at a URL and crawl it" feature — uploads are manual. Bulk file upload is also not yet supported.

No automatic source citations

KB content is injected as context. Replies don't include automatic "Source: [doc name]" markers in the text sent to customers.

One file, one session

Upload files one at a time. The pipeline visualiser shows processing progress in real time so you know when each file is ready.

Editable at any time

Update a chunk or re-embed an entire entry whenever source material changes. The agent picks up new knowledge immediately.

Start building

Upload your first doc.
Have your agent answer
from it in minutes.

7-day free trial. Connect a channel, upload a knowledge base, write your first scenario — and watch the agent answer from your own material.

Ready to build
your first agent?

7-day free trial. Connect your channels, write your first scenarios, and have your agent answering customers within the hour.