Pipeline Steps

Reference for all 13 pipeline steps

Pipeline Steps

Simili Bot’s modular pipeline consists of 13 composable steps.

Step overview

#	Step	Purpose	Speed	Dependencies
1	gatekeeper	Check repo enabled	<1ms	Config
2	command_handler	Parse commands	<10ms	GitHub
3	vectordb_prep	Ensure collection	10-100ms	Qdrant
4	similarity_search	Find related	500ms-1s	Qdrant, Embedder
5	transfer_check	Rule-based route	<100ms	Transfer
6	llm_router	AI routing	2-5s	LLM
7	duplicate_detector	Identify dupes	2-5s	LLM
8	quality_checker	Score issue	1-3s	LLM
9	triage	Suggest labels	1-2s	LLM
10	response_builder	Build comment	<100ms	Results
11	action_executor	Post to GitHub	100-500ms	GitHub API
12	indexer	Add to vector DB	500ms-1s	Qdrant, Embedder
13	pending_actions	Schedule ops	<10ms	State

Detailed steps## Detailed steps

1. gatekeeper

Checks if repository is enabled in configuration.

Input: Issue, Config Output: Skip if disabled Speed: <1ms Dependencies: None

2. command_handler

Processes @simili-bot commands in issue comments.

Input: Comments Output: Commands parsed Speed: <10ms Dependencies: GitHub API

3. vectordb_prep

Ensures Qdrant collection exists, creates if needed.

Input: Config Output: Collection ready Speed: 10-100ms Dependencies: Qdrant

4. similarity_search

Finds related issues using semantic search.

Input: Issue embedding Output: Related issues list Speed: 500ms-1s Dependencies: Qdrant, Embedder

5. transfer_check

Evaluates rule-based routing against configured rules.

Input: Issue metadata Output: Target repository (or empty) Speed: <100ms Dependencies: Transfer matcher

6. llm_router

Uses AI to determine correct repository based on content.

Input: Issue + repo descriptions Output: Routing recommendation Speed: 2-5s Dependencies: LLM (Gemini)

7. duplicate_detector

Analyzes similar issues to identify duplicates.

Input: Similar issues, current issue Output: Duplicate info + confidence score Speed: 2-5s Dependencies: LLM

8. quality_checker

Assesses issue description quality.

Input: Issue content Output: Quality score + improvement suggestions Speed: 1-3s Dependencies: LLM

9. triage

Suggests appropriate labels based on content.

Input: Issue + existing labels Output: Suggested labels with confidence Speed: 1-2s Dependencies: LLM

10. response_builder

Constructs comprehensive analysis comment from results.

Input: All previous step results Output: Formatted comment text Speed: <100ms Dependencies: None (uses previous results)

11. action_executor

Posts to GitHub and applies suggested actions.

Input: Comment text, labels, transfer target Output: GitHub updates Speed: 100-500ms per action Dependencies: GitHub API

12. indexer

Adds or updates issue in vector database for semantic search.

Input: Issue text + metadata Output: Indexed in Qdrant Speed: 500ms-1s Dependencies: Embedder, Qdrant

13. pending_action_scheduler

Schedules actions for later execution.

Input: Pending actions Output: Scheduled operations Speed: <10ms Dependencies: State management

Execution flow

START
  ↓
gatekeeper (Check enabled)
  ↓
command_handler (Parse commands)
  ↓
vectordb_prep (Create collection)
  ↓
similarity_search (Find related)
  ↓
transfer_check (Rule routing)
  ↓
llm_router (AI routing)
  ↓
duplicate_detector (Find dupes)
  ↓
quality_checker (Score)
  ↓
triage (Suggest labels)
  ↓
response_builder (Build message)
  ↓
action_executor (Post actions)
  ↓
indexer (Add to DB)
  ↓
pending_action_scheduler (Schedule)
  ↓
END (Output results)

Each step processes the same Context object, passing data forward.

Step dependencies

Steps can be reordered but dependencies matter:

• similarity_search needs vectordb_prep
• duplicate_detector needs similarity_search
• llm_router optional (needs LLM only)
• response_builder should be near end
• action_executor should be near end

Next steps

Pipeline context

How context flows through steps