feat: Add embedding support to providers#95
Merged
akshaydeo merged 12 commits intomaximhq:mainfrom Jul 3, 2025
Merged
Conversation
Implements embedding functionality with support for Azure, Bedrock, Cohere, Mistral, and OpenAI providers. Adds new EmbeddingRequest type, input/output schemas, provider-specific implementations, fallback support, and HTTP transport integration.
Contributor
Summary by CodeRabbit
Summary by CodeRabbit
WalkthroughThis update introduces unified embedding request support across the Bifrost core, schemas, and multiple provider integrations. It generalizes request handling with a registry pattern, adds embedding-specific methods and schema fields, and implements or stubs embedding logic for OpenAI, Azure, Bedrock, Cohere, and Mistral providers, while standardizing unsupported operation error handling for others. Changes
Sequence Diagram(s)sequenceDiagram
participant Client
participant Bifrost
participant Provider
participant PluginPipeline
Client->>Bifrost: EmbeddingRequest(ctx, req)
Bifrost->>PluginPipeline: RunPreHooks(ctx, req)
alt Short-circuit
PluginPipeline-->>Bifrost: ShortCircuit/Error
Bifrost-->>Client: Error/ShortCircuitResponse
else Continue
Bifrost->>Provider: Embedding(ctx, model, key, input, params)
Provider-->>Bifrost: BifrostResponse/Err
Bifrost->>PluginPipeline: RunPostHooks(ctx, resp, err)
PluginPipeline-->>Bifrost: ModifiedResponse/Error
Bifrost-->>Client: Response/Error
end
Suggested reviewers
Poem
📜 Recent review detailsConfiguration used: CodeRabbit UI 📒 Files selected for processing (1)
🧰 Additional context used🧠 Learnings (2)📓 Common learningscore/providers/bedrock.go (19)🔇 Additional comments (4)
✨ Finishing Touches
🧪 Generate Unit Tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
CodeRabbit Configuration File (
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 20
📜 Review details
Configuration used: CodeRabbit UI
Review profile: ASSERTIVE
Plan: Pro
📒 Files selected for processing (13)
core/bifrost.go(3 hunks)core/providers/anthropic.go(1 hunks)core/providers/azure.go(2 hunks)core/providers/bedrock.go(1 hunks)core/providers/cohere.go(2 hunks)core/providers/mistral.go(2 hunks)core/providers/ollama.go(1 hunks)core/providers/openai.go(3 hunks)core/providers/vertex.go(1 hunks)core/schemas/bifrost.go(3 hunks)core/schemas/provider.go(1 hunks)transports/bifrost-http/main.go(5 hunks)transports/bifrost-http/tracking/plugin.go(1 hunks)
🧰 Additional context used
🧠 Learnings (1)
core/schemas/bifrost.go (1)
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/bifrost.go:46-49
Timestamp: 2025-06-04T09:22:18.123Z
Learning: In core/schemas/bifrost.go, the RequestInput struct uses ChatCompletionInput *[]BifrostMessage (pointer-to-slice) rather than []BifrostMessage to properly represent union type semantics. For text completion requests, ChatCompletionInput should be nil to indicate "no chat payload at all", while for chat completion requests it should be non-nil (even if empty slice). This distinguishes between different request types rather than just empty vs non-empty chat messages.
🪛 golangci-lint (1.64.8)
transports/bifrost-http/main.go
378-378: undefined: schemas.EmbeddingInput
(typecheck)
396-396: unknown field EmbeddingInput in struct literal of type schemas.RequestInput
(typecheck)
405-405: bifrostReq.Params.Dimensions undefined (type *schemas.ModelParameters has no field or method Dimensions)
(typecheck)
406-406: bifrostReq.Params.EncodingFormat undefined (type *schemas.ModelParameters has no field or method EncodingFormat)
(typecheck)
407-407: bifrostReq.Params.User undefined (type *schemas.ModelParameters has no field or method User)
(typecheck)
417-417: client.EmbeddingRequest undefined (type *bifrost.Bifrost has no field or method EmbeddingRequest)
(typecheck)
🔇 Additional comments (4)
transports/bifrost-http/main.go (1)
227-230: Route added successfullyNew
/v1/embeddingsendpoint is wired into the router – nice and symmetric with the existing completion routes.core/schemas/bifrost.go (1)
52-57: Field ordering / zero-value ambiguity
RequestInputnow holds three optional alternatives – good. Note, however, that forEmbeddingInputyou chose a struct pointer whereas for completions you use primitive pointers. The nil/empty-slice semantics are clear, but make sure upstream validation enforces that exactly one of the three fields is non-nil, otherwise ambiguous requests may slip through.core/providers/cohere.go (1)
549-686: API field name may be wrongCohere’s v2 endpoint expects
"embedding_types"? or"embedding_type"(singular). The code sends:"embedding_types": []string{"float"}Please double-check against the latest Cohere docs; an incorrect key will lead to HTTP-400.
core/providers/openai.go (1)
23-37: LGTM! Well-structured embedding response fields.The addition of the
Datafield to handle embedding responses is appropriate. The use ofanytype for theEmbeddingfield correctly accommodates OpenAI's multiple response formats (float arrays and base64-encoded strings).
akshaydeo
reviewed
Jun 19, 2025
| } else { | ||
| result, bifrostError = provider.ChatCompletion(req.Context, req.Model, key, *req.Input.ChatCompletionInput, req.Params) | ||
| } | ||
| } else if req.Type == EmbeddingRequest { |
Contributor
There was a problem hiding this comment.
@Pratham-Mishra04 as a code hygiene - I think nuking this if-else ladder is critical.
We can probably use some factory pattern where
type executor func (provider, message)
type MessageExecutors map[string]executor
const executors = MessageExecutors{<add pre-defined handlers for each message>}
added 6 commits
July 2, 2025 08:55
will need to land and release the core changes first
This commit implements multiple performance, safety, and maintainability improvements across the embedding functionality based on PR feedback: **Performance Optimizations:** - Optimize JSON parsing in Mistral provider using json.RawMessage to avoid double parsing - Improve Azure provider embedding conversion with direct []float64 type assertion - Add input validation for base64 data length in OpenAI provider to prevent runtime panics **Code Quality & Maintainability:** - Add newUnsupportedOperationError helper to reduce code duplication across providers - Make Cohere input_type configurable via ExtraParams for better flexibility - Add clarifying comment for Embedding field JSON tag mapping in schemas - Add early validation for embedding input in EmbeddingRequest function **Safety Improvements:** - Validate base64 decoded data length is divisible by 4 before float32 conversion - Add fail-fast validation for missing embedding input with clear error messages - Improve error handling with more specific and actionable error messages **Files Modified:** - core/bifrost.go: Early validation for embedding input - core/providers/utils.go: Shared helper for unsupported operations - core/providers/cohere.go: Configurable input_type parameter - core/providers/mistral.go: Optimized JSON parsing - core/providers/openai.go: Base64 validation and input validation - core/providers/azure.go: Optimized embedding type conversion - core/providers/bedrock.go: Updated method signatures and token calculation - core/schemas/bifrost.go: Documentation for JSON tag mapping
- Group package-level functions at top (Init, NewPluginPipeline) - Place public API methods before private implementations - Organize private methods by functionality (request handling, infrastructure, initialization) - Add clear section dividers for better navigation - Consolidate try* functions using generic tryRequest implementation - Fix typo: "sepcified" -> "specified" in Bifrost struct comment This improves code readability and follows Go conventions for method organization.
connyay
commented
Jul 2, 2025
…o structure" This reverts commit 837fb0d.
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (2)
core/bifrost.go (2)
468-479: Enhance embedding input validation to prevent unnecessary retries.The current validation only checks for nil but doesn't handle empty strings or zero-length slices. This could cause unnecessary retries when providers return 400 errors for empty input.
Consider checking the actual content:
} else if req.Type == EmbeddingRequest { if req.Input.EmbeddingInput == nil { bifrostError = &schemas.BifrostError{ IsBifrostError: false, Error: schemas.ErrorField{ Message: "input not provided for embedding request", }, } break // Don't retry client errors + } else if embedStr, ok := req.Input.EmbeddingInput.(*string); ok && strings.TrimSpace(*embedStr) == "" { + bifrostError = &schemas.BifrostError{ + IsBifrostError: false, + Error: schemas.ErrorField{ + Message: "embedding input cannot be empty", + }, + } + break // Don't retry client errors + } else if embedSlice, ok := req.Input.EmbeddingInput.(*[]string); ok && len(*embedSlice) == 0 { + bifrostError = &schemas.BifrostError{ + IsBifrostError: false, + Error: schemas.ErrorField{ + Message: "embedding input array cannot be empty", + }, + } + break // Don't retry client errors } else { result, bifrostError = provider.Embedding(req.Context, req.Model, key, req.Input.EmbeddingInput, req.Params) }
468-468: Consider refactoring the if-else ladder using a factory pattern.As @akshaydeo mentioned, this growing if-else ladder impacts code hygiene and maintainability.
📜 Review details
Configuration used: CodeRabbit UI
Review profile: ASSERTIVE
Plan: Pro
📒 Files selected for processing (1)
core/bifrost.go(3 hunks)
🧰 Additional context used
🧠 Learnings (2)
📓 Common learnings
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#84
File: transports/bifrost-http/main.go:2-2
Timestamp: 2025-06-15T16:05:13.489Z
Learning: For the Bifrost project, HTTP transport integration routers for new providers (like Mistral and Ollama) are implemented in separate PRs from the core provider support, following a focused PR strategy.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:89-119
Timestamp: 2025-06-10T11:00:02.875Z
Learning: In the Bifrost OpenAI integration (transports/bifrost-http/integrations/openai/types.go), the convertOpenAIContent function currently only handles the last image URL when multiple images are present in a content array. The user Pratham-Mishra04 has acknowledged this limitation and indicated it's part of a larger architectural issue that will be addressed comprehensively later, rather than with piecemeal fixes.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/provider.go:148-148
Timestamp: 2025-06-04T03:57:50.981Z
Learning: Breaking changes in the Bifrost codebase are managed by first merging and tagging core schema changes, then updating dependent code references in subsequent steps after the core version is released.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/router.go:0-0
Timestamp: 2025-06-09T14:03:34.227Z
Learning: In the Bifrost HTTP transport layer (transports/bifrost-http/integrations/), request validation like checking for empty messages should be handled by the provider rather than at the transport layer. The transport layer should forward requests to Bifrost core/providers for validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/providers/bedrock.go:241-252
Timestamp: 2025-06-04T09:07:20.867Z
Learning: In the Bifrost codebase, when working with AWS Bedrock provider authentication, the preference is to let AWS handle access key validation naturally rather than adding preemptive checks for empty/blank access keys. This allows AWS to provide its own authentication error messages which can be more informative than custom validation errors.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:96-101
Timestamp: 2025-06-15T13:50:41.418Z
Learning: In the Bifrost project, when a provider doesn't support a specific operation (like text completion), the IsBifrostError flag should be set to false to mark it as a provider-level error rather than a Bifrost framework error. This is intentional design for unsupported operations.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#94
File: core/bifrost.go:823-845
Timestamp: 2025-06-18T15:15:51.323Z
Learning: In the Bifrost project, the team prioritizes maintaining consistent error handling patterns over exposing detailed error context. All errors should be wrapped in the standard `BifrostError` structure rather than creating specific error types or exposing richer error details like exit codes or stderr output.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:131-180
Timestamp: 2025-06-10T13:11:37.867Z
Learning: In Anthropic API integration for Bifrost, messages won't contain both image and tool_result content blocks in the same message, so defensive guards against multiple embedded message structs are unnecessary in the content processing loop.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:223-231
Timestamp: 2025-06-10T11:06:06.670Z
Learning: The OpenAI `name` field on messages cannot be preserved when converting to Bifrost format because the `schemas.BifrostMessage` struct in bifrost/core does not support a Name field. This is a known limitation of the Bifrost core schema design.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/schemas/bifrost.go:186-190
Timestamp: 2025-06-15T14:18:32.703Z
Learning: In core/schemas/bifrost.go, the ToolChoice UnmarshalJSON validation intentionally only checks for empty Type fields and lets providers handle validation of specific tool choice values. This architectural decision keeps schema validation focused on structure while allowing provider-specific semantic validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:168-170
Timestamp: 2025-06-15T14:24:49.882Z
Learning: In the Bifrost codebase, performance is prioritized over defensive copying for HTTP service operations. Specifically, shallow slice assignments in provider response handling are acceptable due to object pool reset patterns and JSON unmarshaling behavior that minimize practical data corruption risks.
core/bifrost.go (16)
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/bifrost.go:46-49
Timestamp: 2025-06-04T09:22:18.123Z
Learning: In core/schemas/bifrost.go, the RequestInput struct uses ChatCompletionInput *[]BifrostMessage (pointer-to-slice) rather than []BifrostMessage to properly represent union type semantics. For text completion requests, ChatCompletionInput should be nil to indicate "no chat payload at all", while for chat completion requests it should be non-nil (even if empty slice). This distinguishes between different request types rather than just empty vs non-empty chat messages.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/end_to_end_tool_calling.go:43-45
Timestamp: 2025-06-16T04:13:55.437Z
Learning: In the Bifrost codebase, errors returned from client methods like ChatCompletionRequest are of type BifrostError, not the standard error interface. For testing these errors, use require.Nilf instead of require.NoErrorf since BifrostError doesn't work with the standard error assertion methods.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/router.go:0-0
Timestamp: 2025-06-09T14:03:34.227Z
Learning: In the Bifrost HTTP transport layer (transports/bifrost-http/integrations/), request validation like checking for empty messages should be handled by the provider rather than at the transport layer. The transport layer should forward requests to Bifrost core/providers for validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/provider.go:148-148
Timestamp: 2025-06-04T03:57:50.981Z
Learning: Breaking changes in the Bifrost codebase are managed by first merging and tagging core schema changes, then updating dependent code references in subsequent steps after the core version is released.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/simple_chat.go:39-41
Timestamp: 2025-06-16T04:13:42.755Z
Learning: In the Bifrost codebase, errors returned from methods like ChatCompletionRequest are of type BifrostError (a custom error type) rather than the standard Go error interface. Therefore, require.Nilf should be used for error assertions instead of require.NoErrorf.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#89
File: transports/bifrost-http/integrations/genai/types.go:151-155
Timestamp: 2025-06-16T14:44:42.893Z
Learning: In transports/bifrost-http/integrations/genai/types.go, when SystemInstruction has an empty role, the user prefers to return a validation error rather than setting a default role value. This represents a design preference for strict input validation over silent error correction.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#89
File: transports/bifrost-http/integrations/genai/types.go:151-155
Timestamp: 2025-06-16T14:45:48.563Z
Learning: In transports/bifrost-http/integrations/genai/types.go, when SystemInstruction has an empty role, the user prefers to let the downstream provider (Google GenAI) handle validation and return errors, rather than implementing validation in the bifrost layer. This represents a design preference for delegating validation to the appropriate service rather than duplicating validation logic in the proxy layer.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#131
File: transports/bifrost-http/lib/config.go:35-68
Timestamp: 2025-06-26T07:37:24.385Z
Learning: In the Bifrost project's MCP configuration handling, empty environment variables should be treated as missing/invalid rather than as valid empty values. The current implementation using `os.Getenv(envKey); envValue != ""` to check for non-empty values is the desired behavior.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/schemas/bifrost.go:186-190
Timestamp: 2025-06-15T14:18:32.703Z
Learning: In core/schemas/bifrost.go, the ToolChoice UnmarshalJSON validation intentionally only checks for empty Type fields and lets providers handle validation of specific tool choice values. This architectural decision keeps schema validation focused on structure while allowing provider-specific semantic validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/providers/bedrock.go:241-252
Timestamp: 2025-06-04T09:07:20.867Z
Learning: In the Bifrost codebase, when working with AWS Bedrock provider authentication, the preference is to let AWS handle access key validation naturally rather than adding preemptive checks for empty/blank access keys. This allows AWS to provide its own authentication error messages which can be more informative than custom validation errors.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/complete_end_to_end.go:39-41
Timestamp: 2025-06-16T04:12:05.427Z
Learning: In the Bifrost system, error returns are of type `BifrostError` rather than the standard Go `error` interface. Therefore, use `require.Nilf(t, err, ...)` instead of `require.NoError(t, err)` when checking for errors in Bifrost function calls.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:526-550
Timestamp: 2025-06-04T09:29:46.287Z
Learning: In core/providers/anthropic.go, the content field in formattedMessages is always of type []interface{} because it's explicitly constructed that way upstream in the prepareAnthropicChatRequest function. Defensive type casting for multiple types is not needed since the type is guaranteed by the construction logic.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:0-0
Timestamp: 2025-06-04T05:44:09.141Z
Learning: For the Anthropic provider in core/providers/anthropic.go, it's acceptable to pass potentially malformed messages with invalid roles because Anthropic's API will return suitable error responses for role issues, eliminating the need for additional validation logging.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/utils.go:169-173
Timestamp: 2025-06-09T17:33:52.234Z
Learning: The ChatCompletionRequest method in the Bifrost client follows a contract where the result parameter will never be nil if the error parameter is nil. This means when error checking passes (err == nil), the result is guaranteed to be valid and can be safely used without additional nil checks.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#67
File: transports/bifrost-http/integrations/anthropic/router.go:26-34
Timestamp: 2025-06-10T11:19:29.604Z
Learning: The Generic router in transports/bifrost-http/integrations/utils.go already handles nil pointers from RequestConverter functions. When a RequestConverter returns a nil *schemas.BifrostRequest, the Generic router automatically returns an HTTP 400 error with "Invalid request" message, making additional nil checks in individual router implementations redundant.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#67
File: transports/bifrost-http/integrations/anthropic/router.go:26-34
Timestamp: 2025-06-10T11:19:29.604Z
Learning: The Generic router in transports/bifrost-http/integrations/utils.go already handles nil pointers from RequestConverter functions. When a RequestConverter returns a nil *schemas.BifrostRequest, the Generic router automatically returns an HTTP 400 error, making additional nil checks in individual router implementations redundant.
🔇 Additional comments (2)
core/bifrost.go (2)
25-25: Constant addition follows established pattern.The new
EmbeddingRequestconstant is properly defined and consistent with existing request types.
861-924: Well-implemented embedding request handler with proper validation.The implementation correctly validates inputs early (including the embedding input check that was previously suggested) and follows the established pattern for request handling with fallback support.
* factory pattern for message handlers * combine the try handlers
Contributor
Author
|
I backed out the transports changes as the relationship between core and transports is a bit awkward. Can bring back in a followup change. https://gist.github.com/connyay/f649e4cb53c643ed23d0be1bf7ac8e23 |
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (2)
core/providers/bedrock.go (2)
1222-1225: Make Cohere input_type configurable.The
input_typeis hard-coded to "search_document", but Cohere supports multiple input types ("search_query", "search_document", "classification", "clustering") for different use cases. Consider making this configurable through extra parameters.requestBody := map[string]interface{}{ "texts": input.Texts, - "input_type": "search_document", + "input_type": "search_document", // default } - if params != nil && params.ExtraParams != nil { - maps.Copy(requestBody, params.ExtraParams) - } + if params != nil && params.ExtraParams != nil { + // Allow override of input_type through extra params + if inputType, ok := params.ExtraParams["input_type"]; ok { + requestBody["input_type"] = inputType + delete(params.ExtraParams, "input_type") // Remove to avoid duplication + } + maps.Copy(requestBody, params.ExtraParams) + }
1265-1266: Document token approximation limitations.The token approximation used by
approximateTokenCountis very rough and may significantly differ from actual token usage. Consider adding a comment to document this limitation.// Calculate token usage based on input texts (approximation since Cohere doesn't provide this) + // Note: This is a rough approximation. Actual token count may vary significantly + // depending on the text content and tokenizer used by the model. totalInputTokens := approximateTokenCount(input.Texts)
📜 Review details
Configuration used: CodeRabbit UI
Review profile: ASSERTIVE
Plan: Pro
📒 Files selected for processing (1)
core/providers/bedrock.go(2 hunks)
🧰 Additional context used
🧠 Learnings (2)
📓 Common learnings
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#84
File: transports/bifrost-http/main.go:2-2
Timestamp: 2025-06-15T16:05:13.489Z
Learning: For the Bifrost project, HTTP transport integration routers for new providers (like Mistral and Ollama) are implemented in separate PRs from the core provider support, following a focused PR strategy.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/router.go:0-0
Timestamp: 2025-06-09T14:03:34.227Z
Learning: In the Bifrost HTTP transport layer (transports/bifrost-http/integrations/), request validation like checking for empty messages should be handled by the provider rather than at the transport layer. The transport layer should forward requests to Bifrost core/providers for validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:89-119
Timestamp: 2025-06-10T11:00:02.875Z
Learning: In the Bifrost OpenAI integration (transports/bifrost-http/integrations/openai/types.go), the convertOpenAIContent function currently only handles the last image URL when multiple images are present in a content array. The user Pratham-Mishra04 has acknowledged this limitation and indicated it's part of a larger architectural issue that will be addressed comprehensively later, rather than with piecemeal fixes.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#85
File: core/providers/anthropic.go:150-156
Timestamp: 2025-06-15T16:07:53.140Z
Learning: In the Bifrost codebase, constructor functions are allowed to mutate input ProviderConfig objects in-place (e.g., setting default BaseURL values and trimming trailing slashes). This pattern is acceptable and doesn't need to be flagged as a code review issue.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/provider.go:148-148
Timestamp: 2025-06-04T03:57:50.981Z
Learning: Breaking changes in the Bifrost codebase are managed by first merging and tagging core schema changes, then updating dependent code references in subsequent steps after the core version is released.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/providers/bedrock.go:241-252
Timestamp: 2025-06-04T09:07:20.867Z
Learning: In the Bifrost codebase, when working with AWS Bedrock provider authentication, the preference is to let AWS handle access key validation naturally rather than adding preemptive checks for empty/blank access keys. This allows AWS to provide its own authentication error messages which can be more informative than custom validation errors.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:96-101
Timestamp: 2025-06-15T13:50:41.418Z
Learning: In the Bifrost project, when a provider doesn't support a specific operation (like text completion), the IsBifrostError flag should be set to false to mark it as a provider-level error rather than a Bifrost framework error. This is intentional design for unsupported operations.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#94
File: core/bifrost.go:823-845
Timestamp: 2025-06-18T15:15:51.323Z
Learning: In the Bifrost project, the team prioritizes maintaining consistent error handling patterns over exposing detailed error context. All errors should be wrapped in the standard `BifrostError` structure rather than creating specific error types or exposing richer error details like exit codes or stderr output.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:131-180
Timestamp: 2025-06-10T13:11:37.867Z
Learning: In Anthropic API integration for Bifrost, messages won't contain both image and tool_result content blocks in the same message, so defensive guards against multiple embedded message structs are unnecessary in the content processing loop.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:223-231
Timestamp: 2025-06-10T11:06:06.670Z
Learning: The OpenAI `name` field on messages cannot be preserved when converting to Bifrost format because the `schemas.BifrostMessage` struct in bifrost/core does not support a Name field. This is a known limitation of the Bifrost core schema design.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/schemas/bifrost.go:186-190
Timestamp: 2025-06-15T14:18:32.703Z
Learning: In core/schemas/bifrost.go, the ToolChoice UnmarshalJSON validation intentionally only checks for empty Type fields and lets providers handle validation of specific tool choice values. This architectural decision keeps schema validation focused on structure while allowing provider-specific semantic validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:168-170
Timestamp: 2025-06-15T14:24:49.882Z
Learning: In the Bifrost codebase, performance is prioritized over defensive copying for HTTP service operations. Specifically, shallow slice assignments in provider response handling are acceptable due to object pool reset patterns and JSON unmarshaling behavior that minimize practical data corruption risks.
core/providers/bedrock.go (14)
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/bedrock.go:443-468
Timestamp: 2025-06-04T09:32:15.826Z
Learning: In core/providers/bedrock.go, for tool call result messages (ModelChatMessageRoleTool), the Content field represents the actual tool call output. A tool result message should only be created when msg.Content is non-nil, as there's no semantic meaning to a tool result without output content.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:526-550
Timestamp: 2025-06-04T09:29:46.287Z
Learning: In core/providers/anthropic.go, the content field in formattedMessages is always of type []interface{} because it's explicitly constructed that way upstream in the prepareAnthropicChatRequest function. Defensive type casting for multiple types is not needed since the type is guaranteed by the construction logic.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:168-170
Timestamp: 2025-06-15T14:24:49.882Z
Learning: In the Bifrost codebase, performance is prioritized over defensive copying for HTTP service operations. Specifically, shallow slice assignments in provider response handling are acceptable due to object pool reset patterns and JSON unmarshaling behavior that minimize practical data corruption risks.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:0-0
Timestamp: 2025-06-04T05:44:09.141Z
Learning: For the Anthropic provider in core/providers/anthropic.go, it's acceptable to pass potentially malformed messages with invalid roles because Anthropic's API will return suitable error responses for role issues, eliminating the need for additional validation logging.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:483-494
Timestamp: 2025-06-04T10:04:30.181Z
Learning: In the Anthropic provider (core/providers/anthropic.go), parallel tool calls support including the DisableParallelToolUse flag will be implemented in later commits as it's a relatively new feature that's not commonly used yet. The development approach prioritizes core functionality first.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/cohere.go:327-335
Timestamp: 2025-06-15T13:46:24.869Z
Learning: For Cohere v1 API in core/providers/cohere.go, the tool_choice parameter formatting uses uppercase strings for the "type" field (e.g., "AUTO", "TOOL") and follows a different structure than initially assumed. The current implementation with strings.ToUpper() for the type field is correct for the v1 API.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/cohere.go:327-335
Timestamp: 2025-06-15T13:46:24.869Z
Learning: Cohere v1 API tool_choice parameter accepts only uppercase string values: "REQUIRED" and "NONE". Unlike other providers, it doesn't use structured objects with "type" and "name" fields. The current implementation in core/providers/cohere.go correctly uses strings.ToUpper() to convert ToolChoiceStruct.Type to uppercase format as expected by the API.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/bifrost.go:46-49
Timestamp: 2025-06-04T09:22:18.123Z
Learning: In core/schemas/bifrost.go, the RequestInput struct uses ChatCompletionInput *[]BifrostMessage (pointer-to-slice) rather than []BifrostMessage to properly represent union type semantics. For text completion requests, ChatCompletionInput should be nil to indicate "no chat payload at all", while for chat completion requests it should be non-nil (even if empty slice). This distinguishes between different request types rather than just empty vs non-empty chat messages.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:483-494
Timestamp: 2025-06-04T09:36:27.956Z
Learning: In the Anthropic provider (core/providers/anthropic.go), the user has confirmed through practical experience that the tool_choice parameter should always use an object format with a "type" field (e.g., {"type": "auto"}, {"type": "tool", "name": "function_name"}), even though the official documentation examples sometimes show "auto" as a direct string. The current implementation correctly handles all tool choice types with the object format.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:140-146
Timestamp: 2025-06-10T13:51:52.859Z
Learning: In Bifrost core v1.0.9, ImageContent.Type was a pointer type (*string accessed via bifrost.Ptr), but in v1.0.10 it was changed to a value type (ImageContentType). When reviewing code, check the core version being used to determine the correct assignment pattern.
Learnt from: connyay
PR: maximhq/bifrost#92
File: core/providers/utils.go:60-64
Timestamp: 2025-06-17T14:04:21.104Z
Learning: The bifrost codebase uses string alias types (like `type ContextKey string`) for context keys consistently across multiple packages including plugins/maxim/main.go and transports/bifrost-http/tracking/plugin.go. This pattern should be maintained for consistency rather than switching to the canonical struct{} approach.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:483-494
Timestamp: 2025-06-04T09:36:27.956Z
Learning: The Anthropic API tool_choice parameter always requires an object with a "type" field, even for simple choices like "auto", "any", and "none". The format should be {"type": "auto"} not just "auto" as a string, as confirmed by the official documentation at https://docs.anthropic.com/en/api/messages#tool.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#89
File: transports/bifrost-http/integrations/genai/types.go:22-56
Timestamp: 2025-06-16T14:50:46.859Z
Learning: In the Google GenAI integration at transports/bifrost-http/integrations/genai/types.go, the manual URL-safe base64 decoding implementation (converting - to +, _ to /, and adding padding) is required because base64.RawURLEncoding.DecodeString fails for the specific url encoded bytes format being handled.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/end_to_end_tool_calling.go:43-45
Timestamp: 2025-06-16T04:13:55.437Z
Learning: In the Bifrost codebase, errors returned from client methods like ChatCompletionRequest are of type BifrostError, not the standard error interface. For testing these errors, use require.Nilf instead of require.NoErrorf since BifrostError doesn't work with the standard error assertion methods.
🧬 Code Graph Analysis (1)
core/providers/bedrock.go (3)
core/schemas/bifrost.go (8)
EmbeddingInput(61-63)ModelParameters(89-106)BifrostResponse(299-310)BifrostError(433-440)ErrorField(443-450)LLMUsage(313-319)BifrostResponseExtraFields(414-421)Bedrock(43-43)plugins/mocker/main.go (1)
Usage(137-141)core/schemas/provider.go (1)
Provider(154-163)
🔇 Additional comments (2)
core/providers/bedrock.go (2)
13-13: LGTM: Import addition is correct.The
mapsimport is properly placed and used for themaps.Copyoperation in the Cohere embedding handler.
1095-1110: LGTM: Clean routing implementation.The main
Embeddingmethod provides clean routing based on model prefixes and proper error handling for unsupported models, following existing patterns in the codebase.
AI slop before
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
📜 Review details
Configuration used: CodeRabbit UI
Review profile: ASSERTIVE
Plan: Pro
📒 Files selected for processing (1)
core/providers/bedrock.go(2 hunks)
🧰 Additional context used
🧠 Learnings (2)
📓 Common learnings
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#84
File: transports/bifrost-http/main.go:2-2
Timestamp: 2025-06-15T16:05:13.489Z
Learning: For the Bifrost project, HTTP transport integration routers for new providers (like Mistral and Ollama) are implemented in separate PRs from the core provider support, following a focused PR strategy.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/router.go:0-0
Timestamp: 2025-06-09T14:03:34.227Z
Learning: In the Bifrost HTTP transport layer (transports/bifrost-http/integrations/), request validation like checking for empty messages should be handled by the provider rather than at the transport layer. The transport layer should forward requests to Bifrost core/providers for validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:89-119
Timestamp: 2025-06-10T11:00:02.875Z
Learning: In the Bifrost OpenAI integration (transports/bifrost-http/integrations/openai/types.go), the convertOpenAIContent function currently only handles the last image URL when multiple images are present in a content array. The user Pratham-Mishra04 has acknowledged this limitation and indicated it's part of a larger architectural issue that will be addressed comprehensively later, rather than with piecemeal fixes.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#85
File: core/providers/anthropic.go:150-156
Timestamp: 2025-06-15T16:07:53.140Z
Learning: In the Bifrost codebase, constructor functions are allowed to mutate input ProviderConfig objects in-place (e.g., setting default BaseURL values and trimming trailing slashes). This pattern is acceptable and doesn't need to be flagged as a code review issue.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/provider.go:148-148
Timestamp: 2025-06-04T03:57:50.981Z
Learning: Breaking changes in the Bifrost codebase are managed by first merging and tagging core schema changes, then updating dependent code references in subsequent steps after the core version is released.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/providers/bedrock.go:241-252
Timestamp: 2025-06-04T09:07:20.867Z
Learning: In the Bifrost codebase, when working with AWS Bedrock provider authentication, the preference is to let AWS handle access key validation naturally rather than adding preemptive checks for empty/blank access keys. This allows AWS to provide its own authentication error messages which can be more informative than custom validation errors.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:96-101
Timestamp: 2025-06-15T13:50:41.418Z
Learning: In the Bifrost project, when a provider doesn't support a specific operation (like text completion), the IsBifrostError flag should be set to false to mark it as a provider-level error rather than a Bifrost framework error. This is intentional design for unsupported operations.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#94
File: core/bifrost.go:823-845
Timestamp: 2025-06-18T15:15:51.323Z
Learning: In the Bifrost project, the team prioritizes maintaining consistent error handling patterns over exposing detailed error context. All errors should be wrapped in the standard `BifrostError` structure rather than creating specific error types or exposing richer error details like exit codes or stderr output.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:131-180
Timestamp: 2025-06-10T13:11:37.867Z
Learning: In Anthropic API integration for Bifrost, messages won't contain both image and tool_result content blocks in the same message, so defensive guards against multiple embedded message structs are unnecessary in the content processing loop.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#63
File: transports/bifrost-http/integrations/openai/types.go:223-231
Timestamp: 2025-06-10T11:06:06.670Z
Learning: The OpenAI `name` field on messages cannot be preserved when converting to Bifrost format because the `schemas.BifrostMessage` struct in bifrost/core does not support a Name field. This is a known limitation of the Bifrost core schema design.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/schemas/bifrost.go:186-190
Timestamp: 2025-06-15T14:18:32.703Z
Learning: In core/schemas/bifrost.go, the ToolChoice UnmarshalJSON validation intentionally only checks for empty Type fields and lets providers handle validation of specific tool choice values. This architectural decision keeps schema validation focused on structure while allowing provider-specific semantic validation.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:168-170
Timestamp: 2025-06-15T14:24:49.882Z
Learning: In the Bifrost codebase, performance is prioritized over defensive copying for HTTP service operations. Specifically, shallow slice assignments in provider response handling are acceptable due to object pool reset patterns and JSON unmarshaling behavior that minimize practical data corruption risks.
core/providers/bedrock.go (17)
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/bedrock.go:443-468
Timestamp: 2025-06-04T09:32:15.826Z
Learning: In core/providers/bedrock.go, for tool call result messages (ModelChatMessageRoleTool), the Content field represents the actual tool call output. A tool result message should only be created when msg.Content is non-nil, as there's no semantic meaning to a tool result without output content.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:526-550
Timestamp: 2025-06-04T09:29:46.287Z
Learning: In core/providers/anthropic.go, the content field in formattedMessages is always of type []interface{} because it's explicitly constructed that way upstream in the prepareAnthropicChatRequest function. Defensive type casting for multiple types is not needed since the type is guaranteed by the construction logic.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/mistral.go:168-170
Timestamp: 2025-06-15T14:24:49.882Z
Learning: In the Bifrost codebase, performance is prioritized over defensive copying for HTTP service operations. Specifically, shallow slice assignments in provider response handling are acceptable due to object pool reset patterns and JSON unmarshaling behavior that minimize practical data corruption risks.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:0-0
Timestamp: 2025-06-04T05:44:09.141Z
Learning: For the Anthropic provider in core/providers/anthropic.go, it's acceptable to pass potentially malformed messages with invalid roles because Anthropic's API will return suitable error responses for role issues, eliminating the need for additional validation logging.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:483-494
Timestamp: 2025-06-04T10:04:30.181Z
Learning: In the Anthropic provider (core/providers/anthropic.go), parallel tool calls support including the DisableParallelToolUse flag will be implemented in later commits as it's a relatively new feature that's not commonly used yet. The development approach prioritizes core functionality first.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/cohere.go:327-335
Timestamp: 2025-06-15T13:46:24.869Z
Learning: For Cohere v1 API in core/providers/cohere.go, the tool_choice parameter formatting uses uppercase strings for the "type" field (e.g., "AUTO", "TOOL") and follows a different structure than initially assumed. The current implementation with strings.ToUpper() for the type field is correct for the v1 API.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#83
File: core/providers/cohere.go:327-335
Timestamp: 2025-06-15T13:46:24.869Z
Learning: Cohere v1 API tool_choice parameter accepts only uppercase string values: "REQUIRED" and "NONE". Unlike other providers, it doesn't use structured objects with "type" and "name" fields. The current implementation in core/providers/cohere.go correctly uses strings.ToUpper() to convert ToolChoiceStruct.Type to uppercase format as expected by the API.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#55
File: core/providers/anthropic.go:483-494
Timestamp: 2025-06-04T09:36:27.956Z
Learning: In the Anthropic provider (core/providers/anthropic.go), the user has confirmed through practical experience that the tool_choice parameter should always use an object format with a "type" field (e.g., {"type": "auto"}, {"type": "tool", "name": "function_name"}), even though the official documentation examples sometimes show "auto" as a direct string. The current implementation correctly handles all tool choice types with the object format.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#54
File: core/schemas/bifrost.go:46-49
Timestamp: 2025-06-04T09:22:18.123Z
Learning: In core/schemas/bifrost.go, the RequestInput struct uses ChatCompletionInput *[]BifrostMessage (pointer-to-slice) rather than []BifrostMessage to properly represent union type semantics. For text completion requests, ChatCompletionInput should be nil to indicate "no chat payload at all", while for chat completion requests it should be non-nil (even if empty slice). This distinguishes between different request types rather than just empty vs non-empty chat messages.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:140-146
Timestamp: 2025-06-10T13:51:52.859Z
Learning: In Bifrost core v1.0.9, ImageContent.Type was a pointer type (*string accessed via bifrost.Ptr), but in v1.0.10 it was changed to a value type (ImageContentType). When reviewing code, check the core version being used to determine the correct assignment pattern.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#65
File: transports/bifrost-http/integrations/anthropic/types.go:131-180
Timestamp: 2025-06-10T13:11:37.867Z
Learning: In Anthropic API integration for Bifrost, messages won't contain both image and tool_result content blocks in the same message, so defensive guards against multiple embedded message structs are unnecessary in the content processing loop.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#88
File: core/providers/mistral.go:170-176
Timestamp: 2025-06-16T06:56:55.290Z
Learning: When JSON unmarshaling into pooled structs, slice fields like `Choices []schemas.BifrostResponseChoice` get fresh heap memory allocations from `json.Unmarshal()`. The slice data is not part of the pooled struct's memory, so defensive copying is unnecessary. Resetting pooled structs with `*resp = ResponseType{}` only clears slice headers, not the underlying data.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#67
File: transports/bifrost-http/integrations/anthropic/router.go:26-34
Timestamp: 2025-06-10T11:19:29.604Z
Learning: The Generic router in transports/bifrost-http/integrations/utils.go already handles nil pointers from RequestConverter functions. When a RequestConverter returns a nil *schemas.BifrostRequest, the Generic router automatically returns an HTTP 400 error with "Invalid request" message, making additional nil checks in individual router implementations redundant.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/end_to_end_tool_calling.go:43-45
Timestamp: 2025-06-16T04:13:55.437Z
Learning: In the Bifrost codebase, errors returned from client methods like ChatCompletionRequest are of type BifrostError, not the standard error interface. For testing these errors, use require.Nilf instead of require.NoErrorf since BifrostError doesn't work with the standard error assertion methods.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#67
File: transports/bifrost-http/integrations/anthropic/router.go:26-34
Timestamp: 2025-06-10T11:19:29.604Z
Learning: The Generic router in transports/bifrost-http/integrations/utils.go already handles nil pointers from RequestConverter functions. When a RequestConverter returns a nil *schemas.BifrostRequest, the Generic router automatically returns an HTTP 400 error, making additional nil checks in individual router implementations redundant.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#81
File: tests/core-providers/scenarios/simple_chat.go:39-41
Timestamp: 2025-06-16T04:13:42.755Z
Learning: In the Bifrost codebase, errors returned from methods like ChatCompletionRequest are of type BifrostError (a custom error type) rather than the standard Go error interface. Therefore, require.Nilf should be used for error assertions instead of require.NoErrorf.
Learnt from: Pratham-Mishra04
PR: maximhq/bifrost#89
File: transports/bifrost-http/integrations/genai/types.go:22-56
Timestamp: 2025-06-16T14:50:46.859Z
Learning: In the Google GenAI integration at transports/bifrost-http/integrations/genai/types.go, the manual URL-safe base64 decoding implementation (converting - to +, _ to /, and adding padding) is required because base64.RawURLEncoding.DecodeString fails for the specific url encoded bytes format being handled.
🧬 Code Graph Analysis (1)
core/providers/bedrock.go (3)
core/schemas/bifrost.go (8)
EmbeddingInput(61-63)ModelParameters(89-106)BifrostResponse(299-310)BifrostError(433-440)ErrorField(443-450)LLMUsage(313-319)BifrostResponseExtraFields(414-421)Bedrock(43-43)plugins/mocker/main.go (1)
Usage(137-141)core/schemas/provider.go (1)
Provider(154-163)
🔇 Additional comments (9)
core/providers/bedrock.go (9)
13-13: LGTM - Import addition for maps functionality.The
mapsimport is correctly added to support themaps.Copyfunction used in the Cohere embedding handler.
1096-1110: LGTM - Well-structured embedding method routing.The embedding method properly routes requests based on model prefixes and returns appropriate errors for unsupported models. The error handling correctly sets
IsBifrostError: falsefor unsupported operations, which aligns with the learned pattern from previous reviews.
1114-1126: LGTM - Proper input validation for Titan models.The validation correctly enforces Titan's single-text limitation and provides clear error messages. The error classification is appropriate - using
IsBifrostError: truefor missing input (client error) andIsBifrostError: falsefor multiple texts (provider limitation).
1134-1139: LGTM - Appropriate parameter validation for Titan.The code correctly rejects the
dimensionsparameter for Titan models since they have fixed dimensions. The error message is clear and the error classification is appropriate.
1147-1148: LGTM - Proper URL escaping for AWS signing.The model name is correctly escaped using
url.PathEscapeto ensure proper AWS SIGv4 signing, which is crucial for Bedrock authentication.
1154-1167: Consider optimizing JSON parsing similar to other handlers.The response is parsed directly into
titanRespstruct, which is more efficient than the double-parsing pattern seen in some other parts of the codebase. This approach is preferred.
1191-1196: LGTM - Proper input validation for Cohere models.The validation ensures at least one text is provided and uses appropriate error classification.
1198-1204: LGTM - Cohere request body construction with configurable input_type.The implementation correctly uses
maps.Copyto merge extra parameters, allowing the hardcodedinput_typeto be overridden throughExtraParamsas addressed in previous reviews.
1241-1242: LGTM - Token approximation with documented limitations.The token usage calculation uses the
approximateTokenCountutility function, and based on past reviews, the limitations of this approximation are already documented elsewhere in the codebase.
AI slop before
Collaborator
|
Thank you @connyay your contribution to Bifrost - really solid work. It's been merged in and we hope this won’t be your last. Would love to have you stick around and keep building with us! |
This was referenced Aug 4, 2025
akshaydeo
added a commit
that referenced
this pull request
Nov 17, 2025
feat: Add embedding support to providers
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Implements embedding functionality with support for Azure, Bedrock, Cohere, Mistral, and OpenAI providers.
Adds new EmbeddingRequest type, input/output schemas, and provider-specific implementations.