RAG rewrite

Author: sanchitmonga22

Package.swift

@@ -43,12 +43,6 @@ let useLocalBinaries = false //  Toggle: true for local dev, false for release
 // Updated automatically by CI/CD during releases
 let sdkVersion = "0.19.6"
 
-// RAG binary is only available in local dev mode until the release artifact is published.
-// In remote mode, the RAG xcframework zip + checksum don't exist yet, so including the
-// binary target would block ALL SPM package resolution (not just RAG).
-// Set to true once RABackendRAG-v<version>.zip is published to GitHub releases.
-let ragRemoteBinaryAvailable = false
-
 let package = Package(
     name: "runanywhere-sdks",
     platforms: [
@@ -87,7 +81,7 @@ let package = Package(
             name: "RunAnywhereWhisperKit",
             targets: ["WhisperKitRuntime"]
         ),
-    ] + ragProducts(),
+    ],
     dependencies: [
         .package(url: "https://github.com/apple/swift-crypto.git", from: "3.0.0"),
         .package(url: "https://github.com/Alamofire/Alamofire.git", from: "5.9.0"),
@@ -152,7 +146,8 @@ let package = Package(
                 .product(name: "StableDiffusion", package: "ml-stable-diffusion"),
                 "CRACommons",
                 "RACommonsBinary",
-            ] + ragCoreDependencies(),
+                "RABackendRAGBinary",
+            ],
             path: "sdk/runanywhere-swift/Sources/RunAnywhere",
             exclude: ["CRACommons"],
             swiftSettings: [
@@ -231,62 +226,9 @@ let package = Package(
             path: "sdk/runanywhere-swift/Tests/RunAnywhereTests"
         ),
 
-    ] + ragTargets() + binaryTargets()
+    ] + binaryTargets()
 )
 
-// =============================================================================
-// RAG TARGET HELPERS
-// =============================================================================
-// RAG targets are gated because the remote binary artifact doesn't exist yet.
-// Including a binary target with a placeholder checksum blocks ALL SPM resolution.
-
-/// RAG product (library) — only included when the binary is available
-func ragProducts() -> [Product] {
-    guard useLocalBinaries || ragRemoteBinaryAvailable else { return [] }
-    return [
-        .library(
-            name: "RunAnywhereRAG",
-            targets: ["RAGRuntime"]
-        ),
-    ]
-}
-
-/// RAG dependency for the RunAnywhere core target
-/// NOTE: Core already accesses RAG C headers via CRACommons umbrella (rac_rag.h, rac_rag_pipeline.h).
-/// No additional dependency needed — RAGBackend is only used by RAGRuntime.
-func ragCoreDependencies() -> [Target.Dependency] {
-    return []
-}
-
-/// RAG-related targets (C bridge + Swift runtime)
-func ragTargets() -> [Target] {
-    guard useLocalBinaries || ragRemoteBinaryAvailable else { return [] }
-    return [
-        // C Bridge Module - RAG Backend Headers
-        .target(
-            name: "RAGBackend",
-            dependencies: ["RABackendRAGBinary"],
-            path: "sdk/runanywhere-swift/Sources/RAGRuntime/include",
-            publicHeadersPath: "."
-        ),
-        // RAG Runtime Backend
-        .target(
-            name: "RAGRuntime",
-            dependencies: [
-                "RunAnywhere",
-                "RAGBackend",
-                "ONNXRuntime",
-                "LlamaCPPRuntime",
-            ],
-            path: "sdk/runanywhere-swift/Sources/RAGRuntime",
-            exclude: ["include"],
-            linkerSettings: [
-                .linkedLibrary("c++"),
-            ]
-        ),
-    ]
-}
-
 // =============================================================================
 // BINARY TARGET SELECTION
 // =============================================================================
@@ -369,16 +311,13 @@ func binaryTargets() -> [Target] {
             ),
         ]
 
-        // Only include RAG binary when the release artifact is available
-        if ragRemoteBinaryAvailable {
-            targets.append(
-                .binaryTarget(
-                    name: "RABackendRAGBinary",
-                    url: "https://github.com/RunanywhereAI/runanywhere-sdks/releases/download/v\(sdkVersion)/RABackendRAG-v\(sdkVersion).zip",
-                    checksum: "0000000000000000000000000000000000000000000000000000000000000000" // Replace with actual checksum
-                )
+        targets.append(
+            .binaryTarget(
+                name: "RABackendRAGBinary",
+                url: "https://github.com/RunanywhereAI/runanywhere-sdks/releases/download/v\(sdkVersion)/RABackendRAG-v\(sdkVersion).zip",
+                checksum: "0000000000000000000000000000000000000000000000000000000000000000" // TODO: Replace with actual checksum on release
             )
-        }
+        )
 
         return targets
     }

sdk/runanywhere-commons/CMakeLists.txt

@@ -32,7 +32,7 @@ option(RAC_BUILD_PLATFORM "Build platform backend (Apple Foundation Models, Syst
 option(RAC_BUILD_BACKENDS "Build ML backends (LlamaCPP, ONNX, WhisperCPP, RAG)" OFF)
 option(RAC_BACKEND_LLAMACPP "Build LlamaCPP backend" ON)
 option(RAC_BACKEND_ONNX "Build ONNX backend" ON)
-option(RAC_BACKEND_RAG "Build RAG backend (USearch)" ON)
+option(RAC_BACKEND_RAG "Build RAG pipeline (USearch vector search)" ON)
 # WhisperCPP OFF by default - Sherpa-ONNX (NeMo CTC / Parakeet) is now the primary STT backend
 option(RAC_BACKEND_WHISPERCPP "Build WhisperCPP backend" OFF)
 if(APPLE)
@@ -411,9 +411,9 @@ if(RAC_BUILD_BACKENDS)
         add_subdirectory(src/backends/whispercpp)
     endif()
 
-    if(RAC_BACKEND_RAG AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/src/backends/rag/CMakeLists.txt")
-        message(STATUS "  - RAG backend (USearch)")
-        add_subdirectory(src/backends/rag)
+    if(RAC_BACKEND_RAG AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/src/features/rag/CMakeLists.txt")
+        message(STATUS "  - RAG pipeline (USearch)")
+        add_subdirectory(src/features/rag)
     endif()
 endif()
 

sdk/runanywhere-commons/include/rac/backends/rac_embeddings_onnx.h

@@ -0,0 +1,40 @@
+/**
+ * @file rac_embeddings_onnx.h
+ * @brief RunAnywhere Commons - ONNX Embeddings Backend Public API
+ *
+ * Registration for the ONNX-based embedding provider (sentence-transformer models).
+ * Registers with RAC_CAPABILITY_EMBEDDINGS via the service registry.
+ */
+
+#ifndef RAC_EMBEDDINGS_ONNX_H
+#define RAC_EMBEDDINGS_ONNX_H
+
+#include "rac/core/rac_types.h"
+#include "rac/core/rac_error.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * @brief Register the ONNX embeddings backend
+ *
+ * Registers a service provider for RAC_CAPABILITY_EMBEDDINGS.
+ * Handles .onnx model files with sentence-transformer architecture.
+ *
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_backend_onnx_embeddings_register(void);
+
+/**
+ * @brief Unregister the ONNX embeddings backend
+ *
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_backend_onnx_embeddings_unregister(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* RAC_EMBEDDINGS_ONNX_H */

sdk/runanywhere-commons/include/rac/backends/rac_llm_llamacpp.h

@@ -240,6 +240,69 @@ RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_clear_lora(rac_handle_t handle);
 RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_get_lora_info(rac_handle_t handle,
                                                               char** out_json);
 
+// =============================================================================
+// ADAPTIVE CONTEXT API (for RAG pipelines)
+// =============================================================================
+
+/**
+ * Inject a system prompt into the KV cache at position 0.
+ * Clears existing KV cache first, then decodes the prompt tokens.
+ *
+ * @param handle Service handle (from rac_llm_llamacpp_create)
+ * @param prompt System prompt text
+ * @return RAC_SUCCESS or error code
+ */
+RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_inject_system_prompt(rac_handle_t handle,
+                                                                     const char* prompt);
+
+/**
+ * Append text to the KV cache after current content.
+ * Does not clear existing KV cache — adds at current position.
+ *
+ * @param handle Service handle
+ * @param text Text to append
+ * @return RAC_SUCCESS or error code
+ */
+RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_append_context(rac_handle_t handle,
+                                                               const char* text);
+
+/**
+ * Probe confidence that accumulated context answers a query.
+ * Non-destructive to KV cache — probe tokens are cleaned up.
+ *
+ * @param handle Service handle
+ * @param context Context passage (can be empty if context is in KV cache)
+ * @param query The user question
+ * @param out_confidence Output: confidence in [0.0, 1.0]
+ * @return RAC_SUCCESS or error code
+ */
+RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_probe_confidence(rac_handle_t handle,
+                                                                 const char* context,
+                                                                 const char* query,
+                                                                 float* out_confidence);
+
+/**
+ * Generate response from accumulated KV cache state.
+ * Unlike rac_llm_llamacpp_generate(), does NOT clear the KV cache first.
+ *
+ * @param handle Service handle
+ * @param query Query/suffix to append before generation
+ * @param options Generation options (can be NULL for defaults)
+ * @param out_result Output: Generation result
+ * @return RAC_SUCCESS or error code
+ */
+RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_generate_from_context(
+    rac_handle_t handle, const char* query, const rac_llm_options_t* options,
+    rac_llm_result_t* out_result);
+
+/**
+ * Clear all KV cache state.
+ *
+ * @param handle Service handle
+ * @return RAC_SUCCESS or error code
+ */
+RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_clear_context(rac_handle_t handle);
+
 // =============================================================================
 // BACKEND REGISTRATION
 // =============================================================================

sdk/runanywhere-commons/include/rac/features/llm/rac_llm_service.h

@@ -61,6 +61,30 @@ typedef struct rac_llm_service_ops {
 
     /** Get loaded LoRA adapters info as JSON (optional, NULL if not supported) */
     rac_result_t (*get_lora_info)(void* impl, char** out_json);
+
+    /** Inject system prompt into KV cache at position 0 (optional, NULL if not supported) */
+    rac_result_t (*inject_system_prompt)(void* impl, const char* prompt);
+
+    /** Append text to KV cache after current content (optional, NULL if not supported) */
+    rac_result_t (*append_context)(void* impl, const char* text);
+
+    /**
+     * Probe confidence that accumulated context answers query (optional, NULL if not supported).
+     * Returns confidence in [0.0, 1.0] via out_confidence. Non-destructive to KV cache.
+     */
+    rac_result_t (*probe_confidence)(void* impl, const char* context, const char* query,
+                                     float* out_confidence);
+
+    /**
+     * Generate response from accumulated KV cache state (optional, NULL if not supported).
+     * Unlike generate(), does NOT clear KV cache first.
+     */
+    rac_result_t (*generate_from_context)(void* impl, const char* query,
+                                          const rac_llm_options_t* options,
+                                          rac_llm_result_t* out_result);
+
+    /** Clear all KV cache state (optional, NULL if not supported) */
+    rac_result_t (*clear_context)(void* impl);
 } rac_llm_service_ops_t;
 
 /**
@@ -168,6 +192,78 @@ RAC_API void rac_llm_destroy(rac_handle_t handle);
  */
 RAC_API void rac_llm_result_free(rac_llm_result_t* result);
 
+// =============================================================================
+// ADAPTIVE CONTEXT API - For RAG and similar pipelines
+// =============================================================================
+
+/**
+ * @brief Inject a system prompt into the LLM's KV cache at position 0
+ *
+ * Clears existing KV cache, then seeds with the given prompt.
+ * Optional — returns RAC_ERROR_NOT_SUPPORTED if backend doesn't support it.
+ *
+ * @param handle Service handle
+ * @param prompt System prompt text
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_llm_inject_system_prompt(rac_handle_t handle, const char* prompt);
+
+/**
+ * @brief Append text to the LLM's KV cache after current content
+ *
+ * Does not clear existing KV state — accumulates context incrementally.
+ * Optional — returns RAC_ERROR_NOT_SUPPORTED if backend doesn't support it.
+ *
+ * @param handle Service handle
+ * @param text Text to append
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_llm_append_context(rac_handle_t handle, const char* text);
+
+/**
+ * @brief Probe whether accumulated context answers a query
+ *
+ * Uses logit probing (Yes/No softmax) to estimate confidence.
+ * Non-destructive — probe tokens are removed from KV cache after probing.
+ * Optional — returns RAC_ERROR_NOT_SUPPORTED if backend doesn't support it.
+ *
+ * @param handle Service handle
+ * @param context Context passage (can be empty string if context is already in KV cache)
+ * @param query The user question
+ * @param out_confidence Output: confidence in [0.0, 1.0]
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_llm_probe_confidence(rac_handle_t handle, const char* context,
+                                               const char* query, float* out_confidence);
+
+/**
+ * @brief Generate a response from accumulated KV cache state
+ *
+ * Unlike rac_llm_generate(), this does NOT clear the KV cache first.
+ * Use after inject_system_prompt + append_context to generate from accumulated state.
+ * Optional — returns RAC_ERROR_NOT_SUPPORTED if backend doesn't support it.
+ *
+ * @param handle Service handle
+ * @param query Query/suffix text to append before generation
+ * @param options Generation options (can be NULL for defaults)
+ * @param out_result Output: Generation result
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_llm_generate_from_context(rac_handle_t handle, const char* query,
+                                                    const rac_llm_options_t* options,
+                                                    rac_llm_result_t* out_result);
+
+/**
+ * @brief Clear all KV cache state
+ *
+ * Resets the LLM's context for a fresh adaptive query cycle.
+ * Optional — returns RAC_ERROR_NOT_SUPPORTED if backend doesn't support it.
+ *
+ * @param handle Service handle
+ * @return RAC_SUCCESS or error code
+ */
+RAC_API rac_result_t rac_llm_clear_context(rac_handle_t handle);
+
 #ifdef __cplusplus
 }
 #endif

…ommons/include/backends/rag/ort_guards.h …ns/include/rac/features/rag/ort_guards.hsdk/runanywhere-commons/include/backends/rag/ort_guards.h renamed to sdk/runanywhere-commons/include/rac/features/rag/ort_guards.h sdk/runanywhere-commons/include/backends/rag/ort_guards.h renamed to sdk/runanywhere-commons/include/rac/features/rag/ort_guards.h

@@ -1,8 +1,8 @@
 /**
  * @file rac_rag.h
- * @brief RunAnywhere Commons - RAG Backend Public API
+ * @brief RunAnywhere Commons - RAG Pipeline Public API
  *
- * Registration and control functions for the RAG backend.
+ * Registration and control functions for the RAG pipeline module.
  */
 
 #ifndef RAC_RAG_H
@@ -16,16 +16,17 @@ extern "C" {
 #endif
 
 /**
- * @brief Register the RAG backend module
+ * @brief Register the RAG pipeline module
  *
  * Must be called before using RAG functionality.
+ * Also registers the ONNX embeddings service provider if available.
  *
  * @return RAC_SUCCESS on success, error code otherwise
  */
 RAC_API rac_result_t rac_backend_rag_register(void);
 
 /**
- * @brief Unregister the RAG backend module
+ * @brief Unregister the RAG pipeline module
  *
  * @return RAC_SUCCESS on success, error code otherwise
  */