// SUPPORT RECALL

ASK THE TICKET GRAVEYARD.
GET A CITED ANSWER.

The problem is not “ChatGPT on Jira.” The problem is years of ICIS support tickets and .msg attachments that keyword search cannot recall. SRAG ingests Jira Cloud REST v3 only (v2 search is HTTP 410 on this instance), parses .msg via extract-msg, chunks with overlap, embeds nomic-embed-text at 768 dims through local Ollama, stores srag_chunks.embedding vector(768). Query: cosine nearest chunks, optional MiniMax-Text-01 synthesis via shared.llm.dispatch — model is hardcoded, not an env knob. Changing the embedder means re-ingest. Login for ingest/query/chunks/health UI; public landing. This is not SRM2’s workspace and not Herald. If Ollama or pgvector is down, retrieval is empty — not a hallucinated policy.

Premier pgvector 768 nomic-embed Jira v3
Open ingest Query How it works
768
Embed dim
v3
Jira search
.msg
extract-msg
Ollama
localhost:11434
// HOW IT WORKS

Five steps from a Jira project to a cited paragraph

Ingest is a job. Query is a cosine search. Synthesis is optional and must cite chunks.

01
Connect Jira
JIRA_SERVER / USERNAME / API_KEY. Project default ICISTWO.
02
Ingest + parse
Tickets + .msg attachments. extract-msg required.
03
Chunk + embed
Overlap chunks. nomic-embed-text 768 via Ollama.
04
Ask
/srag/app/query natural language. Cosine over srag_chunks.
05
Optional synthesize
MiniMax-Text-01. Empty retrieval → no invented RCA.
Open SRAG
// QUERY

Plain English, cited chunks — CSS mock

Not Jira’s own search. Not SRM2 Review Board.

meltuc.tech/srag/app/query — CSS mock, not a product screenshot
Askwhy did ICISTWO-… bounce?
Hitscosine chunks
LLMoptional MiniMax
// INGEST

Tickets and .msg, not a crawl of Confluence

v2 search 410 — this app already moved to v3.

meltuc.tech/srag/app — CSS mock, not a product screenshot
APIJira Cloud v3
Mail.msg extract-msg
Storesrag_chunks
// HEALTH

Ollama + pgvector must exist

embedder.py hardcodes the model. Health UI is login-gated.

meltuc.tech/srag/app/health — CSS mock, not a product screenshot
Embednomic-embed-text
ExtCREATE EXTENSION vector
Dim768
// NOT SRM

Recall, not ticket creation

SRM/SRM2 write Jira. SRAG reads history. Do not paste workspace screenshots here.

// WHAT IT REFUSES

No env-swappable embed dim. No v2 search.

Will not invent an answer when the index is empty. Public landing lists no ticket bodies.

// WHO THIS IS FOR

SR managers with years of ICISTWO

People who remember “we saw this in 2023” and cannot find it.

// FEATURES

Ingest, embed, ask, cite

Cards match SRAG README files.

🎫
Jira v3
REST ingest, ICISTWO default.
✉️
.msg parse
extract-msg attachments.
✂️
Chunker
Overlapping windows.
🧬
768-d embed
nomic-embed-text / Ollama.
📐
pgvector
cosine at query time.
💬
NL query
Not keyword JQL only.
🧠
MiniMax synth
Hardcoded model id.
📚
Chunks UI
Browse the index.
❤️
Health UI
Login, not public JSON only.
// GET STARTED

Ingest the history. Ask it later.

Login for /srag/app. This page is not your ticket index.

Open SRAG

Requires a MelTuc account. Create one free.

// SAID SLOWLY

What stays true over time

Support Request RAG System — Jira ticket ingestion, email parsing, natural language search. Health stays cheap. If this page disagrees with the signed-in app, trust the app — the product contract and tests enforce that rule.