feat: Enable web search feature without external services #5

.gitmodules

@@ -0,0 +1,12 @@
+[submodule "dev/rag_api"]
+	path = dev/rag_api
+	url = https://github.com/kalaspuffar/rag_api.git
+	branch = reranker
+[submodule "dev/librechat"]
+	path = dev/librechat
+	url = https://github.com/kalaspuffar/LibreChat.git
+	branch = new/feature/simple_reranker
+[submodule "dev/agents"]
+	path = dev/agents
+	url = https://github.com/kalaspuffar/agents.git
+	branch = simple_reranker

dev/README.md

@@ -0,0 +1,66 @@
+# Development Submodules
+
+Git submodules for local development and testing PRs.
+
+## Setup
+
+### 1. Initialize Submodules
+
+```bash
+git submodule update --init --remote
+```
+
+This checks out the branches specified in `.gitmodules`.
+
+### 1.1. Build Agents Package
+
+Since `agents` is an npm package used by LibreChat, build it before starting:
+
+```bash
+cd dev/agents
+npm install
+npm run build
+cd ../..
+```
+
+### 2. Build and Start
+
+To use local builds from submodules, include the override file:
+
+```bash
+docker compose -f docker-compose.librechat.yml -f docker-compose.librechat.override.yml build
+docker compose -f docker-compose.librechat.yml -f docker-compose.librechat.override.yml up -d
+```
+
+To use published images, omit the override file:
+
+```bash
+docker compose -f docker-compose.librechat.yml build
+docker compose -f docker-compose.librechat.yml up -d
+```
+
+## Testing the Reranker
+
+```bash
+curl -s http://localhost:8000/rerank \
+  -H 'Content-Type: application/json' \
+  -H "Authorization: Bearer $JWT_TOKEN" \
+  -d '{
+    "query": "I love you",
+    "docs": ["I hate you", "I really like you"],
+    "k": 5
+  }'
+```
+
+## Update Submodules
+
+```bash
+git submodule update --remote
+```
+
+## Switch Between Local and Published Images
+
+Since the override file is not automatically loaded, simply include or omit it in your commands:
+
+- **Local builds**: Include `-f docker-compose.librechat.override.yml`
+- **Published images**: Omit the override file flag

docker-compose.librechat.override.yml

@@ -0,0 +1,20 @@
+# Use this file to override the docker-compose.librechat.yml file to enable local builds from dev/ submodules
+# docker-compose will automatically use this file when present
+
+services:
+  api:
+    image: librechat:local
+    build:
+      context: ./dev/librechat
+      dockerfile: Dockerfile
+      target: node
+    volumes:
+      - ./dev/agents/dist:/app/node_modules/@librechat/agents/dist
+
+  rag_api:
+    image: rag_api:local
+    build:
+      context: ./dev/rag_api
+      dockerfile: Dockerfile
+
+

docker-compose.librechat.yml

@@ -79,6 +79,8 @@ services:
   rag_api:
     container_name: rag_api
     image: ghcr.io/danny-avila/librechat-rag-api-dev-lite:latest
+    ports:
+      - "${LIBRECHAT_RAG_PORT:-8000}:8000"
     env_file:
       - .env
     environment:
@@ -87,6 +89,8 @@ services:
       - RAG_OPENAI_API_KEY=${RAG_OPENAI_API_KEY:-${OPENROUTER_API_KEY:-}}
       - RAG_OPENAI_BASEURL=${RAG_OPENAI_BASEURL:-${OPENROUTER_BASE_URL:-}}
       - OPENAI_API_KEY=${OPENAI_API_KEY:-${OPENROUTER_API_KEY:-}}
+      - SIMPLE_RERANKER_MODEL_NAME=${SIMPLE_RERANKER_MODEL_NAME:-mixedbread-ai/mxbai-rerank-large-v2}
+      - SIMPLE_RERANKER_MODEL_TYPE=${SIMPLE_RERANKER_MODEL_TYPE:-cross-encoder}
     restart: always
     depends_on:
       - vectordb

env.example

@@ -38,13 +38,49 @@ LIBRECHAT_MEILI_MASTER_KEY=change-me-meili-master-key
 LIBRECHAT_RAG_PORT=8000
 LIBRECHAT_RAG_API_URL=http://rag_api:8000
 
-
 # RAG API - Embeddings (uses OpenRouter by default via docker-compose)
 # Uncomment and set if you want to use different credentials than OpenRouter
 #RAG_OPENAI_API_KEY=sk-...
 #RAG_OPENAI_BASEURL=https://openrouter.ai/api/v1
 #OPENAI_API_KEY=sk-...
 
+# RAG API - Simple Reranker (local models, not via OpenRouter)
+# Reranking models run locally, not through OpenRouter API
+# See https://github.com/AnswerDotAI/rerankers for available models
+#
+# IMPORTANT: The reranker is called ONCE PER SCRAPED URL during web search.
+# If 5 URLs are scraped, the reranker runs 5 times. Choose models accordingly.
+#
+# Option 1 (high quality but CPU-intensive without GPU):
+SIMPLE_RERANKER_MODEL_NAME=mixedbread-ai/mxbai-rerank-large-v2
+SIMPLE_RERANKER_MODEL_TYPE=cross-encoder
+#
+# RECOMMENDED FOR CPU-ONLY (no GPU): FlashRank models are ONNX-optimized
+# and 5-10x faster on CPU with lower memory footprint. Good multilingual support.
+# Option 2 - FlashRank (fastest, recommended for CPU):
+#SIMPLE_RERANKER_MODEL_NAME=ms-marco-MiniLM-L-12-v2
+#SIMPLE_RERANKER_MODEL_TYPE=flashrank
+#
+# Option 3 - FlashRank default (auto-selects best model):
+#SIMPLE_RERANKER_MODEL_NAME=flashrank
+#SIMPLE_RERANKER_MODEL_TYPE=flashrank
+#
+# Option 4 - Smaller cross-encoder (if you prefer cross-encoder architecture):
+#SIMPLE_RERANKER_MODEL_NAME=cross-encoder/ms-marco-MiniLM-L-6-v2
+#SIMPLE_RERANKER_MODEL_TYPE=cross-encoder
+# Note: L-6-v2 has 22M parameters (very small), L-12-v2 has 33M (small)
+#
+# Option 5 - Multilingual cross-encoder (moderate size, good for German):
+#SIMPLE_RERANKER_MODEL_NAME=BAAI/bge-reranker-base
+#SIMPLE_RERANKER_MODEL_TYPE=cross-encoder
+#
+# PERFORMANCE NOTES:
+# - FlashRank models: Best for CPU-only, ONNX optimized, fast inference
+# - Cross-encoder models: Higher quality but slower on CPU, better with GPU
+# - Large models (like mxbai-rerank-large-v2): Require significant CPU/GPU resources
+# - For German/multilingual: All above models support multiple languages
+# - Model is loaded once at startup and reused for all requests
+
 # LibreChat - Vector Database (PostgreSQL)
 LIBRECHAT_VECTORDB_DB=mydatabase
 LIBRECHAT_VECTORDB_USER=myuser
@@ -58,7 +94,7 @@ GID=1000
 SEARXNG_SECRET_KEY=change-me-searxng-secret
 SEARXNG_BASE_URL=http://localhost:8080/
 SEARXNG_INSTANCE_URL=http://searxng:8080
-SEARXNG_API_KEY=
+SEARXNG_API_KEY=local-selfhost
 
 # Firecrawl - Self-hosted stack
 FIRECRAWL_PORT=3003
@@ -77,7 +113,7 @@ FIRECRAWL_API_KEY=local-selfhost
 FIRECRAWL_VERSION=v2
 FIRECRAWL_NUM_WORKERS=8
 FIRECRAWL_BULL_AUTH_KEY=my-secret-key
-FIRECRAWL_LLM_MODEL=gpt-4o
+FIRECRAWL_LLM_MODEL=mistralai/ministral-8b-2512
 
 # Jina Reranker (temporary - will be replaced with RAG API reranker)
 JINA_API_KEY=

librechat.yaml

@@ -85,8 +85,6 @@ webSearch:
   firecrawlApiKey: "${FIRECRAWL_API_KEY}"
   firecrawlApiUrl: "${FIRECRAWL_API_URL}"
   firecrawlVersion: "${FIRECRAWL_VERSION}"
-  jinaApiKey: "${JINA_API_KEY}"
-  jinaApiUrl: "${JINA_API_URL}"
-  rerankerType: "jina"
+  rerankerType: "simple"
   scraperTimeout: 7500
   safeSearch: 1