embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

embedder

Shared embedding service for ava-brain and mc-brain on pico. One model loaded in RAM, served over the internal Docker network. Auto-unloads after idle to protect server headroom.

What's inside

Base:ollama/ollama:latest
Model:dengcao/Qwen3-Embedding-0.6B:F16 (1024-dim, ~600MB)
API: OpenAI-compatible /v1/embeddings at port 8080
Auto-unload: 5 minutes idle → model unloaded from RAM (OLLAMA_KEEP_ALIVE=5m)
First request after unload: ~5 s warmup, then fast

Endpoints

Path	Auth	Purpose
`POST /v1/embeddings`	none (private network)	OpenAI-compatible embedding API. Used by brain-core's embeddings module.
`GET /api/tags`	none	Healthcheck. Lists loaded models.
`POST /api/pull`	none	Pull a new model into the persistent volume.

Storage layout

Models live in a Docker named volume (embedder-models), NOT on the encrypted LUKS vault. Public model weights aren't sensitive — the encrypted vault is only for PGLite data and audit logs.

embedder-models (Docker volume) → /root/.ollama
└── models/ (qwen3 weights, ~600 MB)

Brain integration

Both ava-brain and mc-brain reach the embedder on the shared Docker network:

BRAIN_EMBED_URL=http://embedder:8080/v1
BRAIN_EMBED_MODEL=dengcao/Qwen3-Embedding-0.6B:F16
BRAIN_EMBED_DIM=1024

(set in each brain's Coolify env)

Reranker (deferred)

brain-core also has a reranker path (BRAIN_RERANKER_URL → /v1/rerank). Ollama doesn't natively expose this endpoint. brain-core/rerank.js already falls back gracefully when the reranker fails — search still works without it, just slightly less precise. Future improvement: add a second container for the rerank model behind /v1/rerank if precision becomes a felt gap.

Local dev

docker compose up --build
curl -X POST http://localhost:8080/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"dengcao/Qwen3-Embedding-0.6B:F16","input":["hello world"]}'| jq '.data[0].embedding | length'# Should print: 1024

Deploy on pico (Coolify)

New Coolify "Application" sourced from this repo
Network: attach to the shared brain-net Docker network
Volume: bind a named volume embedder-models to /root/.ollama
Port: do not expose to host — internal-only on brain-net
Resources: limit to 1.5 CPU / 1.5 GB RAM on Coolify

Memory profile

State	RAM
Idle (model unloaded)	~50 MB (just Ollama runtime)
Loaded (after a request)	~700 MB (Ollama + qwen3 0.6B)
Peak (during inference)	~900 MB

The auto-unload means pico's other services aren't starved during quiet periods.

One-time re-embed (post-cutover)

Because we're switching from llama.cpp+qwen3 (local) to Ollama+Qwen3-Embedding-0.6B (hosted), existing embeddings in pgvector are from a different model variant and won't be directly comparable. After Phase E migration, run once per brain:

# From inside the brain container (or via Coolify exec)
node embed-all.js --force

This re-embeds every embedded row against the new model. Takes ~10 min per brain on cold-start.

Name		Name	Last commit message	Last commit date
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
docker-compose.yml		docker-compose.yml
entrypoint.sh		entrypoint.sh

Name		Name	Last commit message	Last commit date
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
docker-compose.yml		docker-compose.yml
entrypoint.sh		entrypoint.sh

Name		Name	Last commit message	Last commit date
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
docker-compose.yml		docker-compose.yml
entrypoint.sh		entrypoint.sh

Name		Name	Last commit message	Last commit date
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
docker-compose.yml		docker-compose.yml
entrypoint.sh		entrypoint.sh

Name		Name	Last commit message	Last commit date
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
docker-compose.yml		docker-compose.yml
entrypoint.sh		entrypoint.sh

Latest commit

History

Folders and files

Repository files navigation

embedder

What's inside

Endpoints

Storage layout

Brain integration

Reranker (deferred)

Local dev

Deploy on pico (Coolify)

Memory profile

One-time re-embed (post-cutover)

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

embedder

What's inside

Endpoints

Storage layout

Brain integration

Reranker (deferred)

Local dev

Deploy on pico (Coolify)

Memory profile

One-time re-embed (post-cutover)

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

embedder

What's inside

Endpoints

Storage layout

Brain integration

Reranker (deferred)

Local dev

Deploy on pico (Coolify)

Memory profile

One-time re-embed (post-cutover)

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

embedder

What's inside

Endpoints

Storage layout

Brain integration

Reranker (deferred)

Local dev