GrepAI Embedder Configuration

View, change, or troubleshoot the embedding provider and model used by grepai. Handles cascading changes (dimensions, re-indexing, workspace propagation).

Model Reference

Model	Provider	Dims	Speed	Quality	Languages
nomic-embed-text	Ollama	768	Fast	Good	English
nomic-embed-text-v2-moe	Ollama	768	Fast	Better	100+ langs
bge-m3	Ollama	1024	Medium	Excellent	100+ langs
mxbai-embed-large	Ollama	1024	Medium	Better	English
all-minilm	Ollama	384	Very Fast	Basic	English
text-embedding-3-small	OpenAI	1536	Fast (API)	Good	Multi
text-embedding-3-large	OpenAI	3072	Fast (API)	Excellent	Multi

OpenAI pricing: text-embedding-3-small ~$0.02/1M tokens, text-embedding-3-large ~$0.13/1M tokens. Typical project (10k lines) costs ~$0.001.

Workflow

1. Detect Current Configuration

Determine if running in workspace or local mode by checking multiple sources:

Check MCP registration for workspace flag:

claude mcp list 2>/dev/null

Glob: .mcp.json
Read: ~/.claude.json  (look for grepai args with --workspace)

Check local config:

Glob: .grepai/config.yaml

If .grepai/config.yaml exists, read it and extract embedder section.

Check workspace config:

grepai workspace list 2>/dev/null

If workspaces exist, read ~/.grepai/workspace.yaml for embedder details.

Determine active mode:

• If MCP has --workspace {NAME} → workspace mode, config is in ~/.grepai/workspace.yaml
• If .grepai/config.yaml exists and no workspace MCP → local mode
• If both exist → workspace takes precedence for search, local for chunking/ignore

Display current config:

Current Embedder Configuration
─────────────────────────────
Mode:       {workspace: NAME | local}
Config:     {~/.grepai/workspace.yaml | .grepai/config.yaml}
Provider:   {ollama | openai | lmstudio}
Model:      {model name}
Dimensions: {dims}
Endpoint:   {endpoint}

2. Ask: What To Do

Ask via AskUserQuestion:

What would you like to do?

○ Change embedding model (keep same provider)
○ Change embedding provider (e.g. Ollama → OpenAI)
○ View current config (done — already displayed above)
○ Troubleshoot embedding issues

If "View current config" — stop here, already displayed.

3. Change Embedding Model (Same Provider)

3a. Show Available Models

Filter by current provider.

For Ollama:

Available Ollama models:

○ nomic-embed-text — 768 dims, fast, English (Recommended default)
○ nomic-embed-text-v2-moe — 768 dims, fast, 100+ languages
○ bge-m3 — 1024 dims, medium speed, excellent quality, 100+ languages
○ mxbai-embed-large — 1024 dims, medium speed, English

For OpenAI:

Available OpenAI models:

○ text-embedding-3-small — 1536 dims, $0.02/1M tokens (Recommended)
○ text-embedding-3-large — 3072 dims, $0.13/1M tokens, highest quality

Ask via AskUserQuestion with options above.

3b. Check Model Availability (Ollama only)

docker exec {OLLAMA_CONTAINER} ollama list

If selected model not pulled:

Model {MODEL} is not pulled yet. Pull it now?

○ Yes — pull model (may take a few minutes)
○ No — skip, I'll pull it later

If yes:

docker exec {OLLAMA_CONTAINER} ollama pull {MODEL}

To find the Ollama container:

docker ps --filter ancestor=ollama/ollama --format "{{.Names}}" | head -1

3c. Apply Model Change

Look up new dimensions from the model reference table. Apply to the correct config file.

Workspace mode — edit ~/.grepai/workspace.yaml:

Change the embedder section under the active workspace:

embedder:
    model: {NEW_MODEL}
    dimensions: {NEW_DIMS}

Local mode — edit .grepai/config.yaml:

embedder:
  model: {NEW_MODEL}
  dimensions: {NEW_DIMS}

Proceed to step 5 (re-index).

4. Change Embedding Provider

4a. Ask New Provider

Switch to which provider?

○ Ollama (local, free, private)
○ OpenAI (cloud, fast, pay-per-use)
○ LM Studio (local, GUI-based)

4b. Collect Provider-Specific Settings

Switching to Ollama

• Check Ollama is running:

curl -s --max-time 5 http://localhost:11434/api/tags

• Ask for model (show Ollama models from reference table)
• Set endpoint: http://localhost:11434
• Check if model is pulled, offer to pull

Switching to OpenAI

• Ask for API key via AskUserQuestion (or check env $OPENAI_API_KEY)
• Validate key:

python3 -c "
import urllib.request, json
req = urllib.request.Request('https://api.openai.com/v1/models',
    headers={'Authorization': 'Bearer {KEY}'})
resp = urllib.request.urlopen(req)
print(f'Status: {resp.status}')
"

• Ask for model (show OpenAI models from reference table)
• Ask for parallelism:

OpenAI parallelism (concurrent API requests)?

○ 4 (Recommended default — safe for all tiers)
○ 8 (good for Tier 2+)
○ 16 (good for Tier 3+ or high-volume plans)

• Check rate limits to suggest parallelism:

python3 -c "
import urllib.request, json
data = json.dumps({'input': 'test', 'model': '{MODEL}'}).encode()
req = urllib.request.Request('https://api.openai.com/v1/embeddings', data=data,
    headers={'Authorization': 'Bearer {KEY}', 'Content-Type': 'application/json'})
resp = urllib.request.urlopen(req)
for h in resp.headers:
    if 'ratelimit' in h.lower():
        print(f'{h}: {resp.headers[h]}')
"

• Display rate limits and recommend parallelism based on RPM:
  • Under 500 RPM → parallelism 2
  • 500-3000 RPM → parallelism 4-8
  • 3000-10000 RPM → parallelism 8-16
  • Over 10000 RPM → parallelism 16-32

Switching to LM Studio

• Check LM Studio is running:

curl -s --max-time 5 http://127.0.0.1:1234/v1/models

• Set endpoint: http://127.0.0.1:1234
• Ask for model name (user must know which model is loaded)
• Detect dimensions:

curl -s http://127.0.0.1:1234/v1/embeddings \
  -d '{"model": "{MODEL}", "input": ["test"]}' | python3 -c "import sys,json; print(len(json.load(sys.stdin)['data'][0]['embedding']))"

4c. Apply Provider Change

Update the correct config file with all new fields.

Workspace mode — edit ~/.grepai/workspace.yaml:

embedder:
    provider: {NEW_PROVIDER}
    model: {NEW_MODEL}
    endpoint: {NEW_ENDPOINT}
    dimensions: {NEW_DIMS}
    api_key: {KEY}          # OpenAI only
    parallelism: {N}        # OpenAI only, omit for others

Local mode — edit .grepai/config.yaml:

embedder:
  provider: {NEW_PROVIDER}
  model: {NEW_MODEL}
  endpoint: {NEW_ENDPOINT}
  dimensions: {NEW_DIMS}
  api_key: {KEY}            # OpenAI only
  parallelism: {N}          # OpenAI only, omit for others

Remove fields that don't apply to the new provider (e.g., remove api_key when switching to Ollama).

Proceed to step 5 (re-index).

5. Re-Index

CRITICAL: Embeddings from different models are incompatible. The index must be rebuilt after any model or provider change.

Warn the user:

⚠ Changing the embedding model requires a full re-index.
  Old embeddings are incompatible with the new model.
  This will re-process all files in the project/workspace.

Ask via AskUserQuestion:

Re-index now?

○ Yes — clear old index and start re-indexing
○ No — I'll re-index later manually

If yes:

Local mode (GOB backend):

rm -rf .grepai/index.gob .grepai/symbols.gob
grepai watch

Workspace mode (Qdrant):

Delete the old collection and re-index:

# Find collection name (usually workspace_{NAME})
curl -s http://localhost:6333/collections | python3 -c "import sys,json; [print(c['name']) for c in json.load(sys.stdin)['result']['collections']]"

# Delete the old collection
curl -s -X DELETE http://localhost:6333/collections/workspace_{NAME}

# Re-index
grepai watch --workspace {NAME}

Workspace mode (PostgreSQL):

# Truncate the workspace tables (user should confirm)
# Then re-index
grepai watch --workspace {NAME}

If no — print the manual re-index commands for later.

6. Troubleshoot Embedding Issues

If user chose troubleshoot in step 2:

Check 1: Provider connectivity

For Ollama:

curl -s --max-time 5 http://localhost:11434/api/tags

For OpenAI:

python3 -c "
import urllib.request
req = urllib.request.Request('https://api.openai.com/v1/models',
    headers={'Authorization': 'Bearer {KEY}'})
resp = urllib.request.urlopen(req)
print(f'Status: {resp.status}')
"

Check 2: Model availability (Ollama)

docker exec {OLLAMA_CONTAINER} ollama list

Verify the configured model appears in the list.

Check 3: Dimensions match

Test actual embedding dimensions:

# Ollama
curl -s http://localhost:11434/api/embeddings \
  -d '{"model": "{MODEL}", "prompt": "test"}' | python3 -c "import sys,json; print(f'Actual dims: {len(json.load(sys.stdin)[\"embedding\"])}')"

Compare with configured dimensions in config file. Mismatch causes indexing failures.

Check 4: Config file consistency

Read config file and verify:

• model and endpoint are not swapped (common mistake during interactive setup)
• dimensions matches the model reference table
• provider matches the endpoint URL scheme

Check 5: Workspace vs local mismatch

If both workspace and local config exist, verify they use the same model/dimensions. Mixed embeddings in the same store cause search quality issues.

Report all findings with OK/FAIL/WARN indicators and specific fix suggestions.

7. Print Summary

============================================================================
Embedder Configuration Updated
============================================================================

Before: {OLD_PROVIDER} / {OLD_MODEL} ({OLD_DIMS} dims)
After:  {NEW_PROVIDER} / {NEW_MODEL} ({NEW_DIMS} dims)

Config:     {config file path}
Mode:       {workspace: NAME | local}
Re-index:   {Required — started | Required — manual | Not needed}

{IF RE-INDEX STARTED}
Re-indexing in progress. Monitor with:
  grepai watch --workspace {NAME} --status   # workspace
  grepai watch --status                      # local
{END IF}

{IF RE-INDEX MANUAL}
Run these commands to re-index:
  {COMMANDS}
{END IF}
============================================================================