Token-free retrieval engine for past PR review comments. Given code or a concept, it finds the most relevant things reviewers have said before — ranked by similarity AND usefulness.
Two query paths:
| Query | What it matches |
|---|---|
--code <file or diff> |
Structurally similar code patterns (embedding-based) |
--concept "<text>" |
Conceptual matches in comment text (embedding-based) |
search "<text>" |
Keyword matches (FTS, no Python needed) |
query --topic X |
Structured filters by topic, category, reviewer, sentiment |
Quick start
/review-lens setup
Run this first. It checks prerequisites and walks you through setup.
Step 1: Install the binary
curl -sSfL https://raw.githubusercontent.com/futuregerald/review-lens/main/install.sh | bash
Or build from source:
git clone https://github.com/futuregerald/review-lens.git
cd review-lens && CGO_ENABLED=0 go build -o review-lens .
sudo mv review-lens /usr/local/bin/
Verify: review-lens version
Step 2: Set up the Python embedder (optional — needed for similar and embed)
The search, query, stats, and export commands need only the binary.
The similar and embed commands need Python + fastembed for local embeddings.
uv venv --python 3.12 /tmp/rlvenv
VIRTUAL_ENV=/tmp/rlvenv uv pip install fastembed==0.8.0
export PATH=/tmp/rlvenv/bin:$PATH
Verify: review-lens setup --field both
Step 3: Build your review corpus
You need a reviews.db with your team's review comments. Two ways to get one:
Option A: Fetch from GitHub (recommended)
# Fetch review comments for specific reviewers from a repo:
review-lens fetch --reviewer alice --repo myorg/myrepo --db reviews.db
review-lens fetch --reviewer bob --repo myorg/myrepo --db reviews.db
# Classify the comments:
review-lens analyze --db reviews.db
# Embed for semantic search (one-time, ~10 min for 5k comments):
review-lens embed --field both --db reviews.db
Option B: Import from JSON files
review-lens import --dir ./data/ --repo myorg/myrepo --db reviews.db
review-lens analyze --db reviews.db
review-lens embed --field both --db reviews.db
Step 4: Query
# Semantic search — find structurally similar code:
review-lens similar --code app/models/user.rb --k 5 --db reviews.db
# Concept search — find reviews about a pattern:
review-lens similar --concept "blanket rescue hides errors" --k 5 --db reviews.db
# Keyword search (no Python needed):
review-lens search "N+1 eager load" --db reviews.db
# Structured query:
review-lens query --topic error-handling --category bug --db reviews.db
Workflows
/review-lens check
When: Before pushing a PR. Check your diff against past review patterns.
# Check a file:
review-lens similar --code app/interactors/my_feature.rb --k 5 --db reviews.db
# Check a concept:
review-lens similar --concept "adding nil guard on association" --k 5 --db reviews.db
# Pipe a diff:
git diff HEAD~1 | review-lens similar --code - --k 5 --db reviews.db
Look for blocking comments and code suggestions in the results. If senior reviewers have flagged this pattern before, address it proactively.
/review-lens plan
When: After writing an implementation plan, BEFORE writing code.
Extract the plan's key approaches as concept queries:
review-lens similar --concept "blanket rescue in interactors" --k 5 --db reviews.db
review-lens similar --concept "eager loading associations" --k 5 --db reviews.db
Report which patterns have been flagged before and what reviewers said.
/review-lens search
When: Looking for what the team has said about a specific topic.
# Full-text search:
review-lens search "race condition" --db reviews.db --limit 10
# Filter by category and topic:
review-lens query --topic database --category performance --db reviews.db
# Reviewer-specific patterns:
review-lens query --reviewer alice --topic testing --db reviews.db --verbose
/review-lens stats
When: Understanding reviewer patterns and team review culture.
review-lens stats --db reviews.db
review-lens stats --db reviews.db --reviewer alice
Commands
| Command | Needs Python | Description |
|---|---|---|
similar |
Yes | Retrieve similar past reviews by code or concept |
embed |
Yes | Build the vector index (one-time, incremental) |
search |
No | FTS keyword search across comments |
query |
No | Filter by reviewer, category, topic, sentiment |
stats |
No | Reviewer statistics and patterns |
fetch |
No | Fetch review comments from GitHub via gh CLI |
import |
No | Import from JSON files |
analyze |
No | Run heuristic classification on comments |
pattern |
Mixed | Manage the distilled pattern library |
bench |
No | Benchmark embedding models (P@k, MRR) |
export |
No | Export comments as JSON |
threads |
No | View comment threads with replies |
setup |
No | Validate Python env and pre-download models |
How it works
- Two embedding indexes: code (diff_hunk vectors, jina-v2-base-code 768d) and comment (body vectors, bge-small-en-v1.5 384d)
- Model provenance chain: the model used to embed the corpus is stamped in the DB. Queries always use that same model. You can't accidentally mix vectors from different models.
- Usefulness reranking: results are reranked by a composite of similarity and usefulness (blocking comments and code suggestions surface higher). A 0.1+ similarity gap is never overturned by usefulness.
- CGO-free Go binary: SQLite via wazero WASM. No C toolchain needed. Cross-compiles to linux/darwin/windows x amd64/arm64.
JSON output
All query commands support --json for structured output:
review-lens similar --code app/models/user.rb --k 5 --json --db reviews.db
review-lens query --topic testing --json --db reviews.db
review-lens search "N+1" --json --db reviews.db