feat: Add DeepSeek AI provider support

[haosenwang1018]

This PR adds DeepSeek as a new model provider option for Perplexica.

Changes

• Added full LLM implementation with streaming support
• Added embedding model support for document search
• Default models: deepseek-chat and deepseek-reasoner
• OpenAI-compatible API integration
• Configuration fields for API key and base URL

Why DeepSeek?

DeepSeek provides high-quality language models at competitive pricing with strong performance on various benchmarks.

Testing

• Code follows existing provider patterns
• Includes proper TypeScript types
• Error handling implemented

Fixes #XXX (if there's a related issue)

---

Summary by cubic

Added DeepSeek as a new model provider for chat and embeddings using an OpenAI-compatible API. Supports streaming and tool calls with default models deepseek-chat, deepseek-reasoner, and deepseek-embedding.

• New Features

  • Chat LLM with streaming responses and tool call support.
  • Embeddings for documents and queries.
  • Provider config fields: apiKey and baseURL (default https://api.deepseek.com/v1).
  • Registered provider and exposed default models.

• Migration

  • Set DEEPSEEK_API_KEY and optionally DEEPSEEK_BASE_URL.
  • Select DeepSeek models in settings; no code changes needed.

^{Written for commit 16df10e. Summary will update on new commits.}

[cubic-dev-ai]

3 issues found across 4 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/lib/models/providers/deepseek/deepseekLLM.ts">

<violation number="1" location="src/lib/models/providers/deepseek/deepseekLLM.ts:90">
P2: DeepSeek chat-completions payload uses `max_completion_tokens` instead of the documented `max_tokens`, risking ignored limits or request errors.</violation>

<violation number="2" location="src/lib/models/providers/deepseek/deepseekLLM.ts:106">
P2: Unguarded `JSON.parse` of model tool arguments can throw and fail the entire non-stream completion when one tool call contains invalid JSON.</violation>

<violation number="3" location="src/lib/models/providers/deepseek/deepseekLLM.ts:165">
P1: Streaming tool-call reconstruction is mis-indexed: new calls are appended with `push()` but subsequent chunks are merged by `tc.index`, which can corrupt multi-tool outputs.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: Streaming tool-call reconstruction is mis-indexed: new calls are appended with push() but subsequent chunks are merged by tc.index, which can corrupt multi-tool outputs.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/lib/models/providers/deepseek/deepseekLLM.ts, line 165:

<comment>Streaming tool-call reconstruction is mis-indexed: new calls are appended with `push()` but subsequent chunks are merged by `tc.index`, which can corrupt multi-tool outputs.</comment>

<file context>
@@ -0,0 +1,228 @@
+                  id: tc.id!,
+                  arguments: tc.function?.arguments || '',
+                };
+                recievedToolCalls.push(call);
+                return { ...call, arguments: parse(call.arguments || '{}') };
+              } else {
</file context>

[cubic-dev-ai]

P2: DeepSeek chat-completions payload uses max_completion_tokens instead of the documented max_tokens, risking ignored limits or request errors.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/lib/models/providers/deepseek/deepseekLLM.ts, line 90:

<comment>DeepSeek chat-completions payload uses `max_completion_tokens` instead of the documented `max_tokens`, risking ignored limits or request errors.</comment>

<file context>
@@ -0,0 +1,228 @@
+      messages: this.convertToDeepSeekMessages(input.messages),
+      temperature: input.options?.temperature ?? 1.0,
+      top_p: input.options?.topP,
+      max_completion_tokens: input.options?.maxTokens,
+      stop: input.options?.stopSequences,
+      frequency_penalty: input.options?.frequencyPenalty,
</file context>

[cubic-dev-ai]

P2: Unguarded JSON.parse of model tool arguments can throw and fail the entire non-stream completion when one tool call contains invalid JSON.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/lib/models/providers/deepseek/deepseekLLM.ts, line 106:

<comment>Unguarded `JSON.parse` of model tool arguments can throw and fail the entire non-stream completion when one tool call contains invalid JSON.</comment>

<file context>
@@ -0,0 +1,228 @@
+                return {
+                  name: tc.function.name,
+                  id: tc.id,
+                  arguments: JSON.parse(tc.function.arguments),
+                };
+              }
</file context>

[xkonjin]

Quick review pass:

• Main risk area here is auth/session state and stale credential handling.
• I didn’t see targeted regression coverage in the diff; please add or point CI at a focused test for the changed path in deepseekEmbedding.ts, deepseekLLM.ts, index.ts (+1 more).
• Before merge, I’d smoke-test the behavior touched by deepseekEmbedding.ts, deepseekLLM.ts, index.ts (+1 more) with malformed input / retry / rollback cases, since that’s where this class of change usually breaks.